diff --git a/README.md b/README.md
index 2142b9ef7c8f0d8bbea5c0bf87f0ab84f9fd32a9..6d3ea1243e77550d36c3e8b4208dce6c2d2decee 100644
--- a/README.md
+++ b/README.md
@@ -31,7 +31,7 @@ with the properties you want. Six connected layers:
   one for your data; [recipes by problem](https://babu09-meidnet.hf.space/docs/learn/recipes.html).
 * **[Build: Studio](https://babu09-meidnet.hf.space/studio/)** — the live workbench: seven colour-coded blocks
   *Data → Model → Family → Rules → Targets → Search → Candidates*. Change a rule, exclude an element or
-  move a target and watch the effect flow through the later blocks. Bring your own table (CSV / Excel /
+  move a target and see the effect on the later blocks. Bring your own table (CSV / Excel /
   JSON + CIFs), train a model in the browser, run the search, inspect every candidate in 3D, export the
   `meidnet.yaml` that reproduces it on your machine.
 * **[Datasets](https://babu09-meidnet.hf.space/docs/explore/datasets.html)** — datasets that work with
@@ -47,6 +47,12 @@ with the properties you want. Six connected layers:
 **Paper:** A. Babu, R. A. Gouvêa, P. Vandergheynst, G.-M. Rignanese, *npj Computational Materials* (2026),
 [doi:10.1038/s41524-026-02153-3](https://doi.org/10.1038/s41524-026-02153-3).
 
+**Further reading:** A. Babu, N. M. A. Krishnan, *Multimodal and cross-modal learning techniques*, APL Machine
+Learning 4, 030901 (2026), [doi:10.1063/5.0346744](https://doi.org/10.1063/5.0346744) · A. Babu,
+R. Almeida Gouvêa, G.-M. Rignanese, *Toward automated discovery with generative models multimodal learning and
+closed loop workflows in inverse materials design*, Cell Reports Physical Science 7, 103561 (2026),
+[doi:10.1016/j.xcrp.2026.103561](https://doi.org/10.1016/j.xcrp.2026.103561).
+
 This public Space is shared: uploads are capped (8 MB, 1,500 rows, 30 epochs), one training at a time,
 sessions are private per browser tab and removed after an hour. Predicted properties are the model's
 estimates — confirm candidates by DFT or experiment.
diff --git a/meidnet/studio/landing.html b/meidnet/studio/landing.html
index de465015c0435c7f4f7ee3d41a68102bc3f1c650..20175d18f2ac00a6324409125df66dac8501e627 100644
--- a/meidnet/studio/landing.html
+++ b/meidnet/studio/landing.html
@@ -63,6 +63,7 @@ h2.sec{font-size:24px;margin:0 0 6px}.sec-sub{color:var(--muted);margin:0 0 14px
 .how small{display:block;font-weight:500;opacity:.9;font-size:12px}
 .cite{background:var(--card);border:1px solid var(--line);border-radius:14px;padding:18px;margin:0 0 28px}
 .cite pre{background:var(--bg);border:1px solid var(--line);border-radius:10px;padding:12px;overflow:auto;font-size:13px;margin:10px 0 0}
+.reading{margin:0;padding-left:18px;color:var(--muted);font-size:14.5px}.reading li{margin:4px 0}
 .cite button{font:inherit;border:1px solid var(--line);background:var(--card);color:var(--ink);border-radius:9px;padding:6px 12px;cursor:pointer}
 footer{color:var(--faint);font-size:13.5px;padding:10px 0 40px;display:flex;gap:16px;flex-wrap:wrap}
 .theme{font:inherit;border:1px solid var(--line);background:var(--card);color:var(--ink);border-radius:9px;padding:5px 9px;cursor:pointer}
@@ -85,7 +86,7 @@ footer{color:var(--faint);font-size:13.5px;padding:10px 0 40px;display:flex;gap:
     <div>
       <p class="eyebrow">MEIDNet Prism</p>
       <h1>Learn, build and benchmark <span class="prism">multimodal AI</span> for materials discovery</h1>
-      <p class="tag">From what a modality is to a trained model and a reproducible result. MEIDNet is the reference implementation: one space shared by crystal structures and their properties, searched for new materials with the properties you want.</p>
+      <p class="tag">From the concepts of multimodal learning to a trained model and reproducible results. MEIDNet, the reference implementation, learns one space shared by crystal structures and their properties, and searches it for materials with target properties.</p>
       <div class="btns"><a class="btn primary" href="/studio/">▶ Try MEIDNet</a><a class="btn" href="/studio/?panel=data">Use your dataset</a><a class="btn" href="/docs/learn/index.html">Learn multimodality</a></div>
       <p class="pub"><b>Published in npj Computational Materials (2026)</b> · <a href="https://doi.org/10.1038/s41524-026-02153-3">doi 10.1038/s41524-026-02153-3</a> · <a href="https://github.com/ABnano/MEIDNet">GitHub</a> · MIT licence</p>
     </div>
@@ -104,7 +105,7 @@ footer{color:var(--faint);font-size:13.5px;padding:10px 0 40px;display:flex;gap:
   </figure>
 
   <h2 class="sec">Six connected layers</h2>
-  <p class="sec-sub">Learn the ideas, pick a design, build it on your data, and show the evidence. Each layer links to the next.</p>
+  <p class="sec-sub">From concepts to reproducible results: the methods, the choice of architecture, a model trained on your data, and results compared on shared benchmarks.</p>
   <div class="flow" aria-label="The six layers in order">
     <a href="/docs/learn/index.html" style="--accent:var(--c-data)">Learn</a><span>→</span>
     <a href="/docs/learn/architectures.html" style="--accent:var(--c-model)">Architectures</a><span>→</span>
@@ -118,18 +119,18 @@ footer{color:var(--faint);font-size:13.5px;padding:10px 0 40px;display:flex;gap:
       <p><a class="more" href="/docs/learn/index.html">Learn multimodality →</a></p></div>
     <div class="card" style="--accent:var(--c-model)"><h2><small>2 · Architectures</small>Which design fits my data?</h2><p>Early and late fusion, shared latent spaces, cross-attention and contrastive learning, drawn side by side, with an advisor that recommends one.</p>
       <p><a class="more" href="/docs/learn/architectures.html">Architecture Atlas →</a> &nbsp; <a class="more" href="/docs/learn/architectures.html#advisor">Advisor →</a></p></div>
-    <div class="card" style="--accent:var(--c-family)"><h2><small>3 · Build</small>MEIDNet Studio</h2><p>Upload structures and properties, train in the browser, set targets and rules, run the search and inspect every candidate in 3D. No code; one YAML file.</p>
+    <div class="card" style="--accent:var(--c-family)"><h2><small>3 · Build</small>MEIDNet Studio</h2><p>Upload structures and properties, train in the browser, set targets and rules, run the search and inspect every candidate in 3D. Every setting is saved in one YAML file.</p>
       <p><a class="more" href="/studio/">Try MEIDNet →</a> &nbsp; <a class="more" href="/docs/learn/recipes.html">Recipes by problem →</a></p></div>
-    <div class="card" style="--accent:var(--c-rules)"><h2><small>4 · Datasets</small>Data that works</h2><p>Perov-5, the published multimodal benchmark, other datasets with their status, and 31 computed and experimental databases grouped by application.</p>
+    <div class="card" style="--accent:var(--c-rules)"><h2><small>4 · Datasets</small>Datasets and databases</h2><p>Perov-5, the published multimodal benchmark, other datasets used with MEIDNet, and 31 computed and experimental databases grouped by application.</p>
       <p><a class="more" href="/docs/explore/datasets.html">Datasets →</a> &nbsp; <a class="more" href="/docs/explore/databases.html">Databases →</a></p></div>
-    <div class="card" style="--accent:var(--c-targets)"><h2><small>5 · Benchmarks</small>Results with evidence</h2><p>One leaderboard per dataset, an evidence ladder from “generated” to “experimentally validated”, and rows re-run here marked verified.</p>
+    <div class="card" style="--accent:var(--c-targets)"><h2><small>5 · Benchmarks</small>Reproducible results</h2><p>One leaderboard per dataset, an evidence ladder from “generated” to “experimentally validated”, and rows re-run here marked verified.</p>
       <p><a class="more" href="/docs/benchmarks/index.html">Benchmarks →</a></p></div>
-    <div class="card" style="--accent:var(--c-candidates)"><h2><small>6 · Community</small>Share and reproduce</h2><p>Submit a result through GitHub. The maintainer re-runs it and marks it verified when it reproduces; a run that does not agree is shown too.</p>
+    <div class="card" style="--accent:var(--c-candidates)"><h2><small>6 · Community</small>Contribute results</h2><p>Results are submitted through GitHub, re-run by the maintainer and marked as verified when they reproduce. Re-runs that disagree are published as well.</p>
       <p><a class="more" href="/docs/community/contribute.html">Contribute a result →</a></p></div>
   </section>
 
-  <h2 class="sec">Inside Build: the MEIDNet Studio in seven blocks</h2>
-  <p class="sec-sub">The reference implementation. Change one block and the blocks after it update.</p>
+  <h2 class="sec">MEIDNet Studio: from data to candidates in seven blocks</h2>
+  <p class="sec-sub">A change in one block updates every block after it.</p>
   <div class="how" aria-label="The seven blocks of the Studio">
     <div style="--bc:var(--c-data)">Data<small>your materials</small></div><div style="--bc:var(--c-model)">Model<small>shared space</small></div>
     <div style="--bc:var(--c-family)">Family<small>crystal type</small></div><div style="--bc:var(--c-rules)">Rules<small>chemistry</small></div>
@@ -153,8 +154,13 @@ footer{color:var(--faint);font-size:13.5px;padding:10px 0 40px;display:flex;gap:
   year    = {2026},
   doi     = {10.1038/s41524-026-02153-3}
 }</pre>
+    <h3 style="margin:16px 0 4px;font-size:17px">Further reading</h3>
+    <ul class="reading">
+      <li>A. Babu and N. M. A. Krishnan, <i>Multimodal and cross-modal learning techniques</i>, APL Machine Learning <b>4</b>, 030901 (2026). <a href="https://doi.org/10.1063/5.0346744">doi:10.1063/5.0346744</a></li>
+      <li>A. Babu, R. Almeida Gouvêa and G.-M. Rignanese, <i>Toward automated discovery with generative models multimodal learning and closed loop workflows in inverse materials design</i>, Cell Reports Physical Science <b>7</b>, 103561 (2026). <a href="https://doi.org/10.1016/j.xcrp.2026.103561">doi:10.1016/j.xcrp.2026.103561</a></li>
+    </ul>
   </section>
-  <footer><span>MEIDNet Prism: learn, build and benchmark multimodal AI for materials discovery. MEIDNet is its reference implementation.</span><a href="https://github.com/ABnano/MEIDNet">Code</a><a href="/docs/">Documentation</a><a href="/docs/understand/limits.html">Honest scope</a><a href="https://huggingface.co/spaces/Babu09/MEIDNet">Hugging Face Space</a></footer>
+  <footer><span>MEIDNet Prism: learn, build and benchmark multimodal AI for materials discovery. MEIDNet is its reference implementation.</span><a href="https://github.com/ABnano/MEIDNet">Code</a><a href="/docs/">Documentation</a><a href="/docs/understand/limits.html">Scope and roadmap</a><a href="https://huggingface.co/spaces/Babu09/MEIDNet">Hugging Face Space</a></footer>
 </main>
 <script>
 (function(){
diff --git a/meidnet/studio/studio.html b/meidnet/studio/studio.html
index 25f6167c6e93cf5d08fb6d932953ff2e0a648c6c..2e12284a4167145ce28c139cdf48f08ebae50eb8 100644
--- a/meidnet/studio/studio.html
+++ b/meidnet/studio/studio.html
@@ -198,7 +198,7 @@ body>div:empty:not([id]):not([class]){position:absolute;top:0;left:0;width:0;hei
     </div>
   </section>
   <div class="card strip"><div class="strip-scroll"><svg id="pipe" viewBox="0 0 1270 150" aria-label="MEIDNet workflow"></svg></div></div>
-  <div class="chain" id="chain"><span class="lead">Change anything above or in a block — what it does to the blocks after it appears here.</span></div>
+  <div class="chain" id="chain"><span class="lead">Changes made above or in a block are listed here, with their effect on the blocks after it.</span></div>
   <details id="flowlog"><summary>Flow log (0)</summary><ol id="flowlog-list"></ol></details>
   <div class="cols">
     <div id="panel"><div class="card loading-card"><h2>Loading MEIDNet Studio…</h2><p class="muted">Fetching the model and every composition of the perovskite family. This takes a few seconds; on the hosted Studio the very first visit can take up to a minute while the server wakes up.</p><div class="prog"><i class="indeterminate"></i></div></div></div>
@@ -330,7 +330,7 @@ async function boot(){
   if(BLOCKS.some(b=>b.id===qs.get('panel'))) S.panel = qs.get('panel');
   await loadFamily();
   render();
-  chainText(`Ready: ${nfmt(S.ev.rows.length)} compositions, ${nfmt(S.ev.alive.length)} pass every rule. Change anything — its effect on every later block appears here.`);
+  chainText(`Ready: ${nfmt(S.ev.rows.length)} compositions, ${nfmt(S.ev.alive.length)} pass every rule. Changes and their effect on the later blocks are listed here.`);
   if(!STATIC) await resumeJobs();
   if(['space','candidates','data'].includes(qs.get('explore'))) openExplore(qs.get('explore'));
   api('/api/data').then(d=>{ S.data=d; S.dataOwn = !!(S.state.session && S.state.session.checked); statusAll(); updateEdges(); render(); }).catch(()=>{});
@@ -651,7 +651,7 @@ function render(){
   const b=BLOCKS.find(x=>x.id===S.panel); const n=BLOCKS.indexOf(b)+1;
   P.dataset.block=S.panel;
   P.innerHTML = `<div class="card"><div class="eyebrow">Block ${n} of 7 · ${esc(b.name)}</div><h2>${esc(b.name)}</h2><p class="purpose">${esc(b.purpose)}</p>
-    <div class="plain beginner-only"><b>In plain words.</b> ${esc(b.plain)}</div>${ioStrip(S.panel)}${f?f():''}${behind(S.panel)}</div>`;
+    <div class="plain beginner-only"><b>In short.</b> ${esc(b.plain)}</div>${ioStrip(S.panel)}${f?f():''}${behind(S.panel)}</div>`;
   document.querySelectorAll('.node').forEach(nd=>nd.classList.toggle('sel', nd.id==='node-'+S.panel));
   if(!S.ev && S.fam) recompute(null, null); else { statusAll(); updateEdges(); renderLive(); }
   renderLegend(); mountAll3d();
@@ -859,7 +859,7 @@ function trainBox(){
         <tr><td>typical error (MAE)</td>${Object.entries(p.val.mae).map(([k,v])=>`<td>${fmt(v)}${esc(unit(k))}</td>`).join('')}<td>${(100*p.val.retrieval_top1).toFixed(0)}%</td></tr>
         <tr><td>quality</td>${q.map(x=>`<td class="${x.css}">${x.word}</td>`).join('')}<td></td></tr></table>
         <p class="small muted beginner-only">The typical error is compared with how much each property varies in your data: under 25% of that spread is <b>good</b>, under 50% <b>fair</b>, otherwise <b>weak</b>. "Matched correctly" is how often a material's structure and its properties land closest to each other in the model's hidden space (higher is better). A few epochs give a rough model; more epochs or more data usually help.</p>`; }
-    if(st.done&&!st.error) b+=`<div class="verdict good">✅ Model ready — the blocks below now use <b>your</b> model and properties.${st.report?` <a href="/files/${esc(st.report)}" target="_blank">training report</a> · <a href="/files/${esc(st.report.replace('training_report.html','model.pt'))}" download="model.pt">download model.pt</a>`:''}</div>`; }
+    if(st.done&&!st.error) b+=`<div class="verdict good">Model ready. The blocks below now use <b>your</b> model and properties.${st.report?` <a href="/files/${esc(st.report)}" target="_blank">training report</a> · <a href="/files/${esc(st.report.replace('training_report.html','model.pt'))}" download="model.pt">download model.pt</a>`:''}</div>`; }
   return b;
 }
 async function startTrain(){
@@ -912,13 +912,13 @@ function panelFamily(){
   b+=`<h3>Elements per site group <span class="muted small">(click to exclude / include)</span></h3>`;
   for(const [g,grp] of Object.entries(f.groups)){ b+=`<h4>${esc(g)} — ${esc(grp.description)} · ${grp.slots.length} site${grp.slots.length>1?'s':''}</h4><div>`;
     for(const el of grp.sample){ const off=S.ui.exclude.has(el); const ox=(grp.oxidation_states[el]||[]).map(q=>(q>0?'+':'')+q).join('/'); b+=`<span class="chip ${off?'off':''}" onclick="toggleEl(${jsq(el)})" title="charges ${esc(ox)}"><span style="display:inline-block;width:9px;height:9px;border-radius:50%;background:${CPK[el]||'#999'};border:1px solid var(--line);margin-right:5px;vertical-align:-1px"></span>${esc(el)}<small class="muted"> ${esc(ox)}</small></span>`; } b+='</div>'; }
-  b+=`<div class="help beginner-only">Excluding an element removes every composition containing it — watch the Family → Rules numbers change. To change which elements a family allows at all, copy <code>${esc(S.ui.family)}.yaml</code> and edit its <code>groups:</code>.</div>`;
+  b+=`<div class="help beginner-only">Excluding an element removes every composition that contains it; the Family and Rules counts update accordingly. To change which elements a family allows at all, copy <code>${esc(S.ui.family)}.yaml</code> and edit its <code>groups:</code>.</div>`;
   return b;
 }
 
 /* ── Rules ── */
 function panelRules(){
-  const ev=S.ev||evaluateSpace(); let b=`<p class="muted small">Move a limit and watch the count change — the search enforces exactly these windows.</p>`;
+  const ev=S.ev||evaluateSpace(); let b=`<p class="muted small">Moving a limit updates the count; the search applies exactly these windows.</p>`;
   for(const c of S.fam.constraints){ const key=ruleKey(c); const r=S.ui.rules[key]; const d=S.fam.space.rows.map(x=>x.d[key]).filter(Number.isFinite);
     b+=`<div class="rule ${r.enabled?'':'off'}"><div class="hd"><input type="checkbox" ${r.enabled?'checked':''} onchange="toggleRule(${jsq(key)},this.checked)"><b>${esc(ruleTitle(key))}</b><span class="muted small"><b style="color:var(--accent-text)">${nfmt(ev.perRule[key]??0)}</b> of ${nfmt(ev.rows.length)} pass</span></div><p class="small muted">${esc(ruleText(key))}</p>`;
     const num=Object.entries(r.params).filter(([k,v])=>typeof v==='number' && !['cutoff'].includes(k));
@@ -1004,7 +1004,7 @@ function renderLive(){
   L.innerHTML=b;
 }
 function renderLegend(){ const beg=document.body.classList.contains('beginner');
-  $('legend').innerHTML = beg ? `<h4>Glossary</h4><dl class="gloss"><dt>Composition</dt><dd>one element per site of the prototype, e.g. CsPbBr₃.</dd><dt>Rule</dt><dd>a hard check with a measured value and an allowed window.</dd><dt>Target</dt><dd>a property value you want; the model predicts how close each composition gets.</dd><dt>Latent space</dt><dd>the hidden coordinates where structures and properties meet; the search moves through it.</dd><dt>Extrapolation</dt><dd>a target outside the training range — predictions there are guesses.</dd></dl><p class="small muted">Switch on Researcher mode in the header for the configuration as text and the Python behind each block.</p>`
+  $('legend').innerHTML = beg ? `<h4>Glossary</h4><dl class="gloss"><dt>Composition</dt><dd>one element per site of the prototype, e.g. CsPbBr₃.</dd><dt>Rule</dt><dd>a hard check with a measured value and an allowed window.</dd><dt>Target</dt><dd>a property value you want; the model predicts how close each composition gets.</dd><dt>Latent space</dt><dd>the hidden coordinates where structures and properties meet; the search moves through it.</dd><dt>Extrapolation</dt><dd>a target outside the training range; predictions there are less reliable.</dd></dl><p class="small muted">Switch on Researcher mode in the header for the configuration as text and the Python behind each block.</p>`
     : `<h4>How to use</h4><ol class="small" style="padding-left:18px;margin:4px 0"><li><b>Family</b>: pick the crystal type; exclude elements.</li><li><b>Rules</b>: move limits; counts update live.</li><li><b>Targets</b>: set values; see the closest compositions.</li><li><b>Search</b> runs the latent optimisation; <b>Candidates</b> explains each result.</li><li><b>Data/Model</b>: upload your table, train, and design with your own properties.</li></ol>`; }
 
 /* ═══════════════════════════ actions ═══════════════════════════ */
diff --git a/site/404.html b/site/404.html
index c6f003fdb761ac0e176382858724b5790fe8bbdb..95aa07fc31feb2a8c581b05e71f0fe0b44b4dbcf 100644
--- a/site/404.html
+++ b/site/404.html
@@ -686,7 +686,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -885,7 +885,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1594,7 +1594,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/assets/extra.css b/site/assets/extra.css
index a692137433c65cb92bd4fa69fc17d173df5935a0..5498fdd55bb5eea010c7dcafcd93340127b4287a 100644
--- a/site/assets/extra.css
+++ b/site/assets/extra.css
@@ -38,18 +38,7 @@ th[role="columnheader"]:not(.no-sort):hover::after{visibility:visible;opacity:.3
 .bench-table td,.bench-table th{padding:.5em .6em;white-space:nowrap}
 .bench-table td:first-child{white-space:normal;min-width:14em}
 
-/* ── Learn layer: status badges ── */
-.mstatus{display:inline-block;border-radius:999px;padding:0 10px;font-size:.72rem;font-weight:700;border:1px solid;white-space:nowrap;line-height:1.7;vertical-align:middle}
-.mstatus.sup{color:#0a7d0a;border-color:#6fcf6f;background:rgba(10,125,10,.06)}
-.mstatus.core{color:#fff;background:linear-gradient(90deg,#4f46e5,#9333ea);border-color:transparent}
-.mstatus.part{color:#8a5d00;border-color:#eda100}
-.mstatus.inside{color:#1c5cab;border-color:#86b6ef}
-.mstatus.plan{color:#6b6a65;border-color:#b5b4ab;border-style:dashed}
-.mstatus.no{color:#a8261d;border-color:#f0a8a3}
-.mstatus.info{color:#4a3aa7;border-color:#b9b1ef}
-[data-md-color-scheme="slate"] .mstatus.sup{color:#6fcf6f}[data-md-color-scheme="slate"] .mstatus.part{color:#f0c35b}
-[data-md-color-scheme="slate"] .mstatus.inside{color:#86b6ef}[data-md-color-scheme="slate"] .mstatus.plan{color:#c3c2b7}
-[data-md-color-scheme="slate"] .mstatus.no{color:#f0a8a3}[data-md-color-scheme="slate"] .mstatus.info{color:#b9b1ef}
+/* ── Learn layer ── */
 .learn-steps ol{columns:2;column-gap:2em}@media(max-width:700px){.learn-steps ol{columns:1}}
 
 /* ── Architecture Atlas cards and their diagrams ── */
@@ -130,6 +119,9 @@ svg.ad{width:100%;height:auto;display:block;max-width:420px;margin:auto}
 .md-typeset .adv-card h4{margin:.5em 0 .3em;font-size:.95rem}
 .md-typeset .adv-card p,.md-typeset .adv-card li{font-size:.76rem}
 .adv-notes{color:var(--md-default-fg-color--light)}.adv-alt{color:var(--md-default-fg-color--light);font-style:italic}
+.adv-kicker{font-size:.62rem;font-weight:700;letter-spacing:.12em;text-transform:uppercase;color:var(--md-default-fg-color--light);margin-top:2px}
+.md-typeset .adv-card p.adv-where{border-top:1px solid var(--md-default-fg-color--lightest);padding-top:.5em}
+.mm-legend .mm-legend-title{font-weight:700}
 .md-typeset a.adv-btn{display:inline-block;border-radius:9px;padding:5px 12px;font-weight:700;text-decoration:none;border:1px solid var(--md-default-fg-color--lighter);margin:2px 6px 4px 0;font-size:.72rem}
 .md-typeset a.adv-btn.primary{background:#4f46e5;color:#fff;border-color:#4f46e5}
 .md-typeset a.adv-btn.primary:hover{background:#9333ea;border-color:#9333ea}
diff --git a/site/assets/learn.js b/site/assets/learn.js
index 5e8b830424fb5b00af0b412a0a69508411ccb6fa..a4e2c4472da79beaf71739de0ed90b35fed3c94f 100644
--- a/site/assets/learn.js
+++ b/site/assets/learn.js
@@ -4,7 +4,7 @@
   "use strict";
   var STUDIO = "https://babu09-meidnet.hf.space/studio/";
   var STATUS = {
-    sup: "Supported", part: "Partly", inside: "Inside the structure", plan: "Planned",
+    sup: "Available", part: "Partial", inside: "Via the structure", plan: "Planned",
   };
 
   function h(tag, attrs, html) {
@@ -24,7 +24,7 @@
       ex: "Relaxed DFT structures (Materials Project, OQMD, JARVIS, Alexandria) and experimental ones (ICSD, COD). Perov-5 in MEIDNet.",
       enc: "Graph neural networks. MEIDNet uses an E(n)-equivariant graph network (EGNN), so rotating or translating the cell changes nothing; CGCNN is a widely used alternative.",
       combo: "With scalar properties (MEIDNet), with diffraction patterns (structure from a pattern), with text, with images.",
-      meid: "Supported. Up to max_sites atoms per cell (default 20), any elements.",
+      meid: "Available: up to max_sites atoms per cell (default 20), any elements.",
       links: [["Datasets", "../explore/datasets.html"], ["Databases by application", "../explore/databases.html"]],
       refs: [["Satorras et al. 2021 (EGNN)", "https://arxiv.org/abs/2102.09844"], ["Xie and Grossman 2018 (CGCNN)", "https://doi.org/10.1103/PhysRevLett.120.145301"]] },
     { id: "properties", name: "Scalar properties", st: "sup",
@@ -32,7 +32,7 @@
       ex: "In Perov-5: direct band gap (dir_gap, eV) and formation enthalpy (heat_all, eV/atom). Any numeric column of your table.",
       enc: "A small fully connected network (MLP) over the vector of normalised values.",
       combo: "With structures (MEIDNet: design from target properties), with compositions, with spectra.",
-      meid: "Supported. All scalar columns form one vector, read by one MLP; every column becomes a target you can set in the search.",
+      meid: "Available: all scalar columns form one vector, read by one MLP; every column becomes a target that can be set in the search.",
       links: [["Change the target properties", "../recipes/change-targets.html"], ["Add a property", "../recipes/add-property.html"]],
       refs: [["Babu et al. 2026 (MEIDNet)", "https://doi.org/10.1038/s41524-026-02153-3"]] },
     { id: "conditions", name: "Processing conditions", st: "part",
@@ -40,7 +40,7 @@
       ex: "Synthesis and processing tables of experimental groups; growth logs of thin films.",
       enc: "Numbers go into an MLP like properties; categories (atmosphere, method) as one-hot codes.",
       combo: "With composition (early fusion of features), with text (the same information written in a paper).",
-      meid: "Partly. Numeric conditions can be added as extra scalar columns. They are then encoded together with the properties, and the model does not know that they are conditions rather than outcomes.",
+      meid: "Partial: numeric conditions can be added as extra scalar columns. They are encoded together with the properties, so the model does not distinguish conditions from outcomes.",
       links: [["Add a property", "../recipes/add-property.html"]],
       refs: [["Baltrušaitis et al. 2019", "https://doi.org/10.1109/TPAMI.2018.2798607"]] },
     { id: "composition", name: "Composition", st: "inside",
@@ -48,7 +48,7 @@
       ex: "Experimental property tables often give only formulas, for example experimental band gaps and several Matbench tasks.",
       enc: "Networks over the set of elements weighted by their fractions, such as Roost; element descriptors with tree models.",
       combo: "With processing conditions (early fusion), with structures when they are known.",
-      meid: "Inside the structure. The element on each site is part of the structure, so MEIDNet sees the composition, but it cannot train on formulas alone. Formulas that sit on one prototype can be turned into structures on that prototype.",
+      meid: "Via the structure: the element on each site is part of the structure, so the composition is encoded, but formulas alone cannot be used for training. Formulas that sit on one prototype can be turned into structures on that prototype.",
       links: [["Recipe: formula only", "recipes.html#formula-only"]],
       refs: [["Goodall and Lee 2020 (Roost)", "https://doi.org/10.1038/s41467-020-19964-7"]] },
     { id: "text", name: "Text", st: "plan",
@@ -56,7 +56,7 @@
       ex: "Text-mined synthesis recipes: nearly 20,000 procedures extracted from the literature.",
       enc: "Language models: BERT-style encoders or embeddings of large language models.",
       combo: "With structures (contrastive alignment, like images and captions in CLIP), with processing conditions.",
-      meid: "Planned: a text encoder into the shared space. Not implemented.",
+      meid: "Planned: a text encoder that maps into the shared space; not yet available.",
       links: [["Recipe: synthesis text", "recipes.html#synthesis-text"]],
       refs: [["Kononova et al. 2019", "https://doi.org/10.1038/s41597-019-0224-1"], ["Moro, Loh et al. 2025", "https://doi.org/10.1016/j.newton.2025.100016"]] },
     { id: "images", name: "Microscopy images", st: "plan",
@@ -64,7 +64,7 @@
       ex: "Microscopy collections of experimental groups; usually small and rarely paired with structures.",
       enc: "2D convolutional networks or vision transformers.",
       combo: "With processing conditions (how the microstructure formed), with properties measured on the same sample.",
-      meid: "Planned: an image encoder into the shared space. Not implemented.",
+      meid: "Planned: an image encoder that maps into the shared space; not yet available.",
       links: [["Capabilities", "../explore/capabilities.html"]],
       refs: [["Guo et al. 2019", "https://doi.org/10.1109/ACCESS.2019.2916887"]] },
     { id: "spectra", name: "Spectra and DOS", st: "plan",
@@ -72,7 +72,7 @@
       ex: "Computed DOS in the Materials Project and JARVIS; measured spectra.",
       enc: "1D convolutional networks or transformers over the binned curve.",
       combo: "With structures (which sites shape which features), with properties (a band gap read from the DOS).",
-      meid: "Planned: as vector modalities with an encoder each. Not implemented.",
+      meid: "Planned: vector modalities with one encoder each; not yet available.",
       links: [["Scope and roadmap", "../understand/limits.html"]],
       refs: [["Tsai et al. 2019 (attention across sequences)", "https://doi.org/10.18653/v1/P19-1656"]] },
     { id: "xrd", name: "Diffraction (XRD)", st: "plan",
@@ -80,7 +80,7 @@
       ex: "Patterns simulated from any structure database (for example with pymatgen); measured powder patterns.",
       enc: "1D convolutional networks; transformers over peak lists.",
       combo: "With structures (identify a structure from its pattern), with compositions.",
-      meid: "Planned: binned XRD as a vector modality, the first item of the roadmap. Not implemented.",
+      meid: "Planned: binned XRD patterns as a vector modality, the first item of the roadmap; not yet available.",
       links: [["Recipe: XRD to structure", "recipes.html#xrd-structure"]],
       refs: [["Park et al. 2017", "https://doi.org/10.1107/S205225251700714X"]] },
   ];
@@ -98,12 +98,12 @@
   };
 
   function detailHtml(m) {
-    return '<h4>' + m.name + ' <span class="mstatus ' + m.st + '">' + STATUS[m.st] + "</span></h4>" +
+    return "<h4>" + m.name + "</h4>" +
       '<table class="mm-facts"><tr><th>What it is</th><td>' + m.what + "</td></tr>" +
       "<tr><th>Examples</th><td>" + m.ex + "</td></tr>" +
       "<tr><th>Networks that read it</th><td>" + m.enc + "</td></tr>" +
       "<tr><th>Combined with</th><td>" + m.combo + "</td></tr>" +
-      "<tr><th>In MEIDNet</th><td>" + m.meid + "</td></tr>" +
+      "<tr><th>MEIDNet</th><td>" + m.meid + "</td></tr>" +
       "<tr><th>Go further</th><td>" + links(m.links) + "</td></tr>" +
       "<tr><th>References</th><td>" + links(m.refs) + "</td></tr></table>";
   }
@@ -134,8 +134,8 @@
         '<text class="st" x="' + pos[i][0].toFixed(1) + '" y="' + (pos[i][1] + 14).toFixed(1) + '" text-anchor="middle">' + STATUS[m.st] + "</text></g>");
     });
     svg.push("</svg>");
-    var legend = '<div class="mm-legend"><span><i class="sw sup"></i>Supported in MEIDNet</span><span><i class="sw part"></i>Partly</span>' +
-      '<span><i class="sw inside"></i>Inside the structure</span><span><i class="sw plan"></i>Planned</span></div>';
+    var legend = '<div class="mm-legend"><span class="mm-legend-title">MEIDNet:</span><span><i class="sw sup"></i>available</span>' +
+      '<span><i class="sw part"></i>partial</span><span><i class="sw inside"></i>via the structure</span><span><i class="sw plan"></i>planned</span></div>';
     var list = '<div class="mm-list" role="group" aria-label="Modalities"><button type="button" data-id="core" class="core-btn">Shared representation</button>' +
       MODS.map(function (m) { return '<button type="button" data-id="' + m.id + '" class="' + m.st + '">' + m.name + "<small>" + STATUS[m.st] + "</small></button>"; }).join("") + "</div>";
     root.innerHTML = legend + svg.join("") + list + '<div class="mm-detail" aria-live="polite"></div>';
@@ -225,7 +225,12 @@
   var GOALS = [["design", "Design new materials from target properties"], ["predict", "Predict properties"],
     ["retrieve", "Find materials that match a profile, across modalities"], ["identify", "Identify a structure or phase from a measurement"],
     ["combine", "Combine all my data into one prediction"]];
-  var LABEL = { sup: "Supported in MEIDNet", part: "Partly in MEIDNet", plan: "Planned in MEIDNet", no: "Not in MEIDNet: use another tool", info: "First step" };
+  var WHERE = {
+    sup: "MEIDNet: the Studio, or the meidnet command line.",
+    part: "MEIDNet provides the structure and property encoders and their fusion; a prediction head for another quantity is not included.",
+    plan: "Planned for MEIDNet and not yet available; the single-modality models named above can be used today.",
+    no: "Outside the scope of MEIDNet; the models named above are the usual choice.",
+  };
   var EXTRA_NAMES = { xrd: "diffraction patterns", spectra: "spectra", images: "images", text: "text" };
 
   function advise(has, goal) {
@@ -240,11 +245,11 @@
     var extra = ["xrd", "spectra", "images", "text"].filter(function (k) { return has[k]; }).map(function (k) { return EXTRA_NAMES[k]; });
     var n = Object.keys(has).filter(function (k) { return has[k]; }).length;
     function meidnetNotes() {
-      if (extra.length) notes.push("MEIDNet uses structures and scalar properties today. Your " + extra.join(", ") +
-        " would each need an encoder of their own, which is planned; train on the structures and properties now.");
-      if (K) notes.push("Processing conditions can be added as extra scalar columns. They are encoded with the properties, and the model does not know they are conditions rather than outcomes.");
+      if (extra.length) notes.push("MEIDNet currently uses crystal structures and scalar properties. The " + extra.join(", ") +
+        " would each need a dedicated encoder, which is planned; the structures and properties can be used now.");
+      if (K) notes.push("Processing conditions can be added as extra scalar columns. They are encoded with the properties, so the model does not distinguish conditions from outcomes.");
     }
-    if (!n) return { st: "info", title: "Tick the data you have", why: "The recommendation depends on which modalities describe your materials." };
+    if (!n) return { st: "info", title: "Select the data you have", why: "The recommendation depends on which modalities describe your materials." };
 
     if (goal === "design") {
       if (S && P) {
@@ -260,6 +265,7 @@
       }
       if (P && C && !S) return { st: "no", arch: "early-fusion", title: "A composition model in a screening loop, or structures built on a prototype",
         why: "MEIDNet needs a structure for every row. If all your formulas sit on one prototype (for example ABX₃), build those structures on the prototype and use MEIDNet. Otherwise train a composition model such as Roost and use it to screen candidate formulas.",
+        where: "Roost or another composition model; MEIDNet requires a crystal structure for each material.",
         links: [["Recipe: formula only", "recipes.html#formula-only", 1], ["Roost (Goodall and Lee 2020)", "https://doi.org/10.1038/s41467-020-19964-7"]] };
       if (S && !P) return { st: "info", arch: "shared-latent", title: "First add property values to your structures",
         why: "Design from properties needs examples of structures with their property values. Compute them (DFT) or take them from a database, then train the shared space.",
@@ -268,27 +274,29 @@
         why: "The same design as MEIDNet, with an encoder suited to your modality in place of the structure encoder. Design then returns " + extra.join(" or ") + " rather than structures, which is useful only if those can be turned into materials.",
         links: [["Shared latent space", "#shared-latent", 1], ["Scope and roadmap", "../understand/limits.html"]] };
       return { st: "info", title: "Add a description of each material besides its properties",
-        why: "Design needs to output something you can make: structures (MEIDNet) or at least formulas. Tick crystal structures or formulas.",
+        why: "Design needs to output something that can be made: structures (MEIDNet) or at least formulas. Select crystal structures or formulas.",
         links: [["Datasets", "../explore/datasets.html", 1]] };
     }
 
     if (goal === "predict") {
       if (!P) return { st: "info", title: "Add the property values you want to predict",
-        why: "A model learns to predict a property from examples where the property is known. Tick scalar properties if you have them.",
+        why: "A model learns to predict a property from examples where the property is known. Select scalar properties if you have them.",
         links: [["Databases by application", "../explore/databases.html", 1]] };
       if (S) {
         if (extra.length) notes.push("To add your " + extra.join(", ") + ", join one encoder per modality in a shared space (intermediate fusion), or use late fusion if some materials lack them.");
         return { st: "sup", arch: "shared-latent", title: "A graph network on the structure (MEIDNet's structure encoder does this)",
           why: "For structure to property alone, a single-modality graph network such as CGCNN is the simplest strong choice. MEIDNet's structure encoder does the same and scores every composition of a family instantly, and it adds the reverse direction if you later want to design.",
+          where: "MEIDNet's structure encoder (the Studio or meidnet space), or a graph network such as CGCNN.",
           links: [["Recipe: screen a family", "recipes.html#screen-family", 1], ["Open the Studio", STUDIO], ["CGCNN (Xie and Grossman 2018)", "https://doi.org/10.1103/PhysRevLett.120.145301"]] };
       }
       if (C && !extra.length) return { st: "no", arch: K ? "early-fusion" : null, title: "A composition network such as Roost" + (K ? ", with the conditions joined by early fusion" : ""),
         why: "Without structures, the formula is the description. Roost learns from the formula alone; numeric conditions can be concatenated to its features.",
+        where: "Roost or another composition model; MEIDNet requires a crystal structure for each material.",
         links: [["Recipe: formula only", "recipes.html#formula-only", 1], ["Roost (Goodall and Lee 2020)", "https://doi.org/10.1038/s41467-020-19964-7"]] };
       if (extra.length === 1 && !C && !K) {
         var one = { "diffraction patterns": "a 1D convolutional network on the binned pattern", spectra: "a 1D convolutional network on the binned spectrum",
           images: "a 2D convolutional network or a vision transformer", text: "a fine-tuned language model" }[extra[0]];
-        return { st: "no", title: "One modality: " + one, why: "With a single input modality, a network suited to its shape is the right start. Multimodal designs help once a second modality is paired with it.",
+        return { st: "no", title: "One modality: " + one, why: "With a single input modality, a network suited to its shape is the natural starting point. Multimodal designs become useful once a second modality is paired with it.",
           links: [["Modalities and their networks", "index.html#what-is-a-modality", 1]] };
       }
       return { st: "no", arch: "late-fusion", title: "Late fusion, or early fusion if every material has every modality",
@@ -310,32 +318,33 @@
         why: "The same loss that pairs pictures with captions pairs your " + extra.join(" and ") + " with structures, given enough pairs. MEIDNet plans these encoders; they are not implemented.",
         links: [["Contrastive learning", "#contrastive", 1], ["Recipe: synthesis text", "recipes.html#synthesis-text"]] };
       return { st: "info", title: "Retrieval needs pairs of two modalities",
-        why: "Tick at least two modalities that describe the same materials, for example crystal structures and scalar properties." };
+        why: "Select at least two modalities that describe the same materials, for example crystal structures and scalar properties." };
     }
 
     if (goal === "identify") {
       if (X || Sp) {
         if (S) notes.push("With structures as well, a contrastive model can retrieve the matching structure instead of a class label; simulate the patterns from the structures to get as many pairs as you need.");
         return { st: "plan", arch: "contrastive", title: "A 1D convolutional classifier on the pattern; contrastive retrieval as the multimodal version",
-          why: "A pattern or spectrum is a fingerprint. A classifier needs labels (crystal system, space group, phase); a contrastive pattern-structure model needs only pairs. In MEIDNet the pattern encoder is planned.",
+          why: "A pattern or spectrum is a fingerprint. A classifier needs labels (crystal system, space group, phase); a contrastive pattern-structure model needs only pairs.",
+          where: "1D convolutional classifiers can be used today; the contrastive version needs the pattern encoder planned for MEIDNet.",
           links: [["Recipe: XRD to structure", "recipes.html#xrd-structure", 1], ["Park et al. 2017", "https://doi.org/10.1107/S205225251700714X"]] };
       }
       if (I) return { st: "no", title: "An image classifier: a convolutional network or a vision transformer",
         why: "Phases and microstructures can be classified from micrographs given labelled examples. This is a single-modality task outside MEIDNet.",
         links: [["Microscopy images on the map", "index.html#the-multimodality-map", 1]] };
       return { st: "info", title: "Identification starts from a measurement",
-        why: "Tick diffraction patterns, spectra or images: the measurement you want to identify a structure or phase from." };
+        why: "Select diffraction patterns, spectra or images: the measurement to identify a structure or phase from." };
     }
 
     // combine
-    if (n < 2) return { st: "info", title: "Combining needs at least two modalities", why: "Tick every kind of data you have for the same materials." };
+    if (n < 2) return { st: "info", title: "Combining needs at least two modalities", why: "Select every kind of data available for the same materials." };
     var tabular = ["structure", "xrd", "spectra", "images", "text"].every(function (k) { return !has[k]; });
     notes.push("If some materials lack a modality, late fusion (one model per modality, predictions averaged) is the robust choice.");
     if (tabular) return { st: "no", arch: "early-fusion", title: "Early fusion: concatenate the features into one model",
       why: "Properties, formulas and conditions all become fixed-length numbers, so concatenating them is a strong and cheap baseline.",
       links: [["Early fusion", "#early-fusion", 1], ["Late fusion", "#late-fusion"]] };
     if (T || (n >= 3 && (X || Sp || I))) return { st: "no", arch: "cross-attention", title: "Cross-attention fusion, with a shared latent space as the cheaper start",
-      why: "Text, spectra or images next to structures have parts that interact in detail; attention learns which parts matter to which. It needs much data. A shared latent space (one encoder per modality, aligned and averaged) is the cheaper first step.",
+      why: "Text, spectra or images next to structures have parts that interact in detail; attention learns which parts matter to which. It needs much data. A shared latent space (one encoder per modality, aligned and averaged) is a less expensive first step.",
       links: [["Cross-attention", "#cross-attention", 1], ["Shared latent space", "#shared-latent"]] };
     return { st: S && P && !extra.length ? "part" : "no", arch: "shared-latent", title: "Intermediate fusion: one encoder per modality into a shared space, then a prediction head",
       why: "Modalities of different shapes each get the encoder that suits them, and their latents are joined. MEIDNet does this for structures and scalar properties (aligned, then averaged); a separate prediction head for another quantity is not part of it.",
@@ -343,11 +352,13 @@
   }
 
   function renderAdvice(r, notes) {
-    var b = '<div class="adv-card ' + r.st + '"><span class="mstatus ' + r.st + '">' + LABEL[r.st] + "</span><h4>" + r.title + "</h4>" +
+    var b = '<div class="adv-card ' + r.st + '"><div class="adv-kicker">' + (r.st === "info" ? "Next step" : "Recommendation") + "</div><h4>" + r.title + "</h4>" +
       "<p><b>Why.</b> " + r.why + "</p>";
     if (r.steps) b += "<ol>" + r.steps.map(function (s) { return "<li>" + s + "</li>"; }).join("") + "</ol>";
     if (notes.length) b += '<ul class="adv-notes">' + notes.map(function (s) { return "<li>" + s + "</li>"; }).join("") + "</ul>";
     if (r.alt) b += '<p class="adv-alt">' + r.alt + "</p>";
+    var where = r.where || WHERE[r.st];
+    if (where) b += '<p class="adv-where"><b>Where to run it.</b> ' + where + "</p>";
     var ls = (r.links || []).slice();
     if (r.arch && !ls.some(function (l) { return l[1] === "#" + r.arch; })) ls.push(["This architecture in the Atlas", "#" + r.arch]);
     if (ls.length) b += '<p class="adv-links">' + ls.map(function (l) {
diff --git a/site/benchmarks/carbon24.html b/site/benchmarks/carbon24.html
index ac6565d94792a8818e2b74bbffc3843d46824c92..9b7446a1568b178ad95413a9c350c527b780426f 100644
--- a/site/benchmarks/carbon24.html
+++ b/site/benchmarks/carbon24.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1622,7 +1622,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/benchmarks/index.html b/site/benchmarks/index.html
index 232e6847b1a19a3500c59c19478ce073fdc70c52..3dc1249d4fc8df064de7e5a179357b4a3dc12747 100644
--- a/site/benchmarks/index.html
+++ b/site/benchmarks/index.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1703,7 +1703,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/benchmarks/mp20.html b/site/benchmarks/mp20.html
index c4807e7cea2e1974dea7ec2e73b9454f64fa1f7e..a526e179c5b83b55a6e2a6117b2076b952bc2ddb 100644
--- a/site/benchmarks/mp20.html
+++ b/site/benchmarks/mp20.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1622,7 +1622,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/benchmarks/perov5.html b/site/benchmarks/perov5.html
index d14ccf395c55183f6166003f94ec19f7f1b4ec38..b3d7c69449d0021ed4268e1de9d2eff2c98a5210 100644
--- a/site/benchmarks/perov5.html
+++ b/site/benchmarks/perov5.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1703,7 +1703,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2370,7 +2370,7 @@ evidence level (generated → ML filtered → MLIP validated → DFT validated 
 <li>Best <code>mae_dir_gap</code>: <strong>Shipped checkpoint re-evaluated on the training split (this code)</strong> (2.78 eV), meidnet verified.</li>
 <li>Best <code>r2_dir_gap</code>: <strong>Shipped checkpoint re-evaluated on the training split (this code)</strong> (-29.8), meidnet verified.</li>
 <li>Evidence levels present: generated, MLIP validated, DFT validated. 2 of 5 rows have been reproduced here.</li>
-<li>2 row(s) are quoted from the paper; a verified re-run of the same checkpoint, where it exists, is the row to trust for the exact numbers of this code version.</li>
+<li>2 row(s) are quoted from the paper; a verified re-run of the same checkpoint, where it exists, gives the exact numbers for this code version.</li>
 <li><strong>A re-run of <a href="results/paper-representation.html">MEIDNet (paper): alignment of structure and property latents on Perov-5</a> here did not agree</strong> (<code>cosine_matched</code> 0.97 → 0.771; <code>l2_matched</code> 0.24 → 0.676). The row keeps its status; see its page for the full comparison.</li>
 </ul>
 
diff --git a/site/benchmarks/results/paper-inverse-design.html b/site/benchmarks/results/paper-inverse-design.html
index 0d5ee07aad6be2a1347f370bf6e68a0fb7488ab9..d4410997c7ac66458e72eaa6db0119cdbc7f9b59 100644
--- a/site/benchmarks/results/paper-inverse-design.html
+++ b/site/benchmarks/results/paper-inverse-design.html
@@ -693,7 +693,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -892,7 +892,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1601,7 +1601,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/benchmarks/results/paper-representation.html b/site/benchmarks/results/paper-representation.html
index 052a3b350d046c131c29b5d339b354d09080fd5e..1df30daae95641a0e1d9c34ef12450a04a2938c5 100644
--- a/site/benchmarks/results/paper-representation.html
+++ b/site/benchmarks/results/paper-representation.html
@@ -693,7 +693,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -892,7 +892,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1601,7 +1601,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/benchmarks/results/paper-results-mace-screen.html b/site/benchmarks/results/paper-results-mace-screen.html
index 16b78cd200fc3229202c9b09c2e3f3b81ba967d1..03608f156a98fd5e121180dbd44dd5e977174582 100644
--- a/site/benchmarks/results/paper-results-mace-screen.html
+++ b/site/benchmarks/results/paper-results-mace-screen.html
@@ -693,7 +693,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -892,7 +892,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1601,7 +1601,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/benchmarks/results/shipped-checkpoint-train-split.html b/site/benchmarks/results/shipped-checkpoint-train-split.html
index 2cbe1f626ef6cb048f8a156c347d90c480fe816c..8775cf6a568f8abbfe404d066970ea3285fc3132 100644
--- a/site/benchmarks/results/shipped-checkpoint-train-split.html
+++ b/site/benchmarks/results/shipped-checkpoint-train-split.html
@@ -693,7 +693,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -892,7 +892,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1601,7 +1601,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/benchmarks/results/shipped-checkpoint-val-split.html b/site/benchmarks/results/shipped-checkpoint-val-split.html
index f3d01af7e82056df912ab1730ec6a51c57d1032f..6a36a25cea1e71ad9a74c8f13be2100a12bd3184 100644
--- a/site/benchmarks/results/shipped-checkpoint-val-split.html
+++ b/site/benchmarks/results/shipped-checkpoint-val-split.html
@@ -693,7 +693,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -892,7 +892,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1601,7 +1601,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/community/contribute.html b/site/community/contribute.html
index 656a1e02178e4d5ddbfdc4a6f693b01fa250e827..6968caa48fc02b764b518f5138185cc6caef6ec7 100644
--- a/site/community/contribute.html
+++ b/site/community/contribute.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1225,10 +1225,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#the-five-things-people-get-wrong" class="md-nav__link">
+  <a href="#common-mistakes" class="md-nav__link">
     <span class="md-ellipsis">
       
-        The five things people get wrong
+        Common mistakes
       
     </span>
   </a>
@@ -1692,7 +1692,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2121,10 +2121,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#the-five-things-people-get-wrong" class="md-nav__link">
+  <a href="#common-mistakes" class="md-nav__link">
     <span class="md-ellipsis">
       
-        The five things people get wrong
+        Common mistakes
       
     </span>
   </a>
@@ -2215,7 +2215,7 @@ version:</p>
 </tr>
 </tbody>
 </table>
-<h2 id="the-five-things-people-get-wrong">The five things people get wrong<a class="headerlink" href="#the-five-things-people-get-wrong" title="Permanent link">&para;</a></h2>
+<h2 id="common-mistakes">Common mistakes<a class="headerlink" href="#common-mistakes" title="Permanent link">&para;</a></h2>
 <ol>
 <li><strong>Mixing categories</strong> in one file. A representation result (MAE, R², retrieval) and a design result (SUN rate)
    are two files.</li>
@@ -2236,12 +2236,12 @@ version:</p>
 </ol>
 <p><strong>Verified</strong> means <em>reproduced from the submitted configuration with this code</em>. It does not mean the material is
 real: that is what the validation level says. A re-run that does <strong>not</strong> agree is kept too
-(<code>benchmarks/verified/&lt;id&gt;.attempt.json</code>) and shown on the row's page next to the submitted numbers, so that a
-disagreement is visible rather than silent — this applies to the paper's own rows as much as to anyone else's.</p>
+(<code>benchmarks/verified/&lt;id&gt;.attempt.json</code>) and shown on the row's page next to the submitted numbers, so that
+disagreements are visible. This applies to the paper's own rows as well.</p>
 <p>Representation rows that are reproduced here also get their per-material predictions saved
 (<code>benchmarks/verified/&lt;id&gt;.predictions.csv</code>), which the pages turn into parity plots and a latent-agreement histogram.</p>
 <p>If you prefer not to open a pull request, the issue template <em>Benchmark submission</em> asks for the same fields and the
-maintainer converts it. Submissions are reviewed by the author of MEIDNet; there is no promised response time.</p>
+maintainer converts it. Submissions are reviewed by the author of MEIDNet.</p>
 
 
 
diff --git a/site/community/index.html b/site/community/index.html
index 573035418c838d9e60a5dced85b3b992d21b8009..d50ebea102dd89274b7d4fa6aeaecf70e1bf1cc4 100644
--- a/site/community/index.html
+++ b/site/community/index.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1622,7 +1622,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/examples/double-perovskite.html b/site/examples/double-perovskite.html
index 0d6e8e9423c8cedee531c2cb8c94d5d82732d3aa..a8dbb8a579cfb95a394a2c455460e12150e5834a 100644
--- a/site/examples/double-perovskite.html
+++ b/site/examples/double-perovskite.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1612,7 +1612,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/examples/perov5.html b/site/examples/perov5.html
index bab25d95576262298624b11a92b1d8c8688bfdee..56201b3c7cf7112752c4770b56f914ac428ef0c7 100644
--- a/site/examples/perov5.html
+++ b/site/examples/perov5.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1612,7 +1612,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/explore/capabilities.html b/site/explore/capabilities.html
index aa301ef5dbaefdfd92a84c9cf4873c043fa73532..64d7d6e0ef5ed319ae5ab64a28dcd39222e133d7 100644
--- a/site/explore/capabilities.html
+++ b/site/explore/capabilities.html
@@ -27,7 +27,7 @@
     
     
       
-        <title>Capabilities (what works today) - MEIDNet Prism</title>
+        <title>Capabilities - MEIDNet Prism</title>
       
     
     
@@ -80,7 +80,7 @@
     <div data-md-component="skip">
       
         
-        <a href="#capabilities-what-works-today" class="md-skip">
+        <a href="#capabilities" class="md-skip">
           Skip to content
         </a>
       
@@ -113,7 +113,7 @@
         <div class="md-header__topic" data-md-component="header-topic">
           <span class="md-ellipsis">
             
-              Capabilities (what works today)
+              Capabilities
             
           </span>
         </div>
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -912,7 +912,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -930,7 +930,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1736,7 +1736,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2250,8 +2250,8 @@
   
 
 
-<h1 id="capabilities-what-works-today">Capabilities: what works today<a class="headerlink" href="#capabilities-what-works-today" title="Permanent link">&para;</a></h1>
-<p>Generated from the code, so it cannot promise more than the code does. <strong>Supported</strong> = in this release; <strong>Community-tested</strong> = a benchmark result reproduced here names it; <strong>Planned</strong> = designed for, not implemented (<a href="../understand/limits.html">roadmap</a>).</p>
+<h1 id="capabilities">Capabilities<a class="headerlink" href="#capabilities" title="Permanent link">&para;</a></h1>
+<p>This page is generated from the code. <strong>Available</strong>: in this release. <strong>Community-tested</strong>: used by a benchmark result reproduced here. <strong>Planned</strong>: on the <a href="../understand/limits.html">roadmap</a>, not yet implemented.</p>
 <h2 id="modalities">Modalities<a class="headerlink" href="#modalities" title="Permanent link">&para;</a></h2>
 <table>
 <thead>
@@ -2300,7 +2300,7 @@
 </tbody>
 </table>
 <h2 id="fusion">Fusion<a class="headerlink" href="#fusion" title="Permanent link">&para;</a></h2>
-<p>One scheme: <strong>early fusion</strong> - the structure latent and the property latent of a material are averaged into the joint latent (<code>meidnet/model.py</code>), after a contrastive alignment whose weight ramps up over <code>training.contrastive_warmup_epochs</code> (the curriculum of the paper). There is no late-fusion option; a selector would be a fiction.</p>
+<p>The structure latent and the property latent of a material are averaged into the joint latent (<code>meidnet/model.py</code>); the paper calls this early fusion. The two latents are first aligned by a contrastive loss whose weight increases over <code>training.contrastive_warmup_epochs</code> (the curriculum of the paper). Other fusion schemes are compared in the <a href="../learn/architectures.html">Architecture Atlas</a>.</p>
 <h2 id="input-formats">Input formats<a class="headerlink" href="#input-formats" title="Permanent link">&para;</a></h2>
 <table>
 <thead>
diff --git a/site/explore/databases.html b/site/explore/databases.html
index d925efd80c5d516aab66f29bd50a872e845ed1d2..75f3d672971fbf506555c7d0bc59e712340654ed 100644
--- a/site/explore/databases.html
+++ b/site/explore/databases.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -1071,7 +1071,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1780,7 +1780,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/explore/datasets.html b/site/explore/datasets.html
index 66797104821cd614ab592e1142838e0c29083091..08ee24b2fd5fd2ecc3bb9717ba94bcf100df3295 100644
--- a/site/explore/datasets.html
+++ b/site/explore/datasets.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -911,10 +911,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#datasets-we-would-like-to-see-benchmarked" class="md-nav__link">
+  <a href="#next-datasets" class="md-nav__link">
     <span class="md-ellipsis">
       
-        Datasets we would like to see benchmarked
+        Next datasets
       
     </span>
   </a>
@@ -972,7 +972,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1681,7 +1681,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2110,10 +2110,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#datasets-we-would-like-to-see-benchmarked" class="md-nav__link">
+  <a href="#next-datasets" class="md-nav__link">
     <span class="md-ellipsis">
       
-        Datasets we would like to see benchmarked
+        Next datasets
       
     </span>
   </a>
@@ -2141,9 +2141,8 @@
 
 
 <h1 id="datasets">Datasets<a class="headerlink" href="#datasets" title="Permanent link">&para;</a></h1>
-<p>Datasets that work with MEIDNet, and what has been done with each. <strong>One card per dataset, never one score
-across datasets</strong>: the tasks, sizes and properties differ, so results are only compared within a dataset
-(see <a href="../benchmarks/index.html">Benchmarks</a>).</p>
+<p>Datasets used with MEIDNet and what has been done with each. Tasks, sizes and properties differ from one dataset to
+the next, so results are compared within a dataset only (see <a href="../benchmarks/index.html">Benchmarks</a>).</p>
 <table>
 <thead>
 <tr>
@@ -2151,7 +2150,7 @@ across datasets</strong>: the tasks, sizes and properties differ, so results are
 <th>structures</th>
 <th>properties used</th>
 <th>size</th>
-<th>with MEIDNet</th>
+<th>available here</th>
 </tr>
 </thead>
 <tbody>
@@ -2160,35 +2159,35 @@ across datasets</strong>: the tasks, sizes and properties differ, so results are
 <td>cubic ABX₃ perovskites, 5 atoms per cell</td>
 <td>direct band gap, formation enthalpy</td>
 <td>18,928</td>
-<td><span class="mstatus sup">trained model, configuration, tutorial, benchmark</span></td>
+<td>pretrained model, configuration, tutorial, benchmark results</td>
 </tr>
 <tr>
 <td><a href="#mp-20">MP-20</a></td>
 <td>Materials Project, up to 20 atoms</td>
 <td>none aligned in the paper</td>
 <td>about 45,000</td>
-<td><span class="mstatus plan">structure representation, published</span></td>
+<td>structure-representation results in the paper</td>
 </tr>
 <tr>
 <td><a href="#carbon-24">Carbon-24</a></td>
 <td>carbon allotropes, up to 24 atoms</td>
 <td>none aligned in the paper</td>
 <td>about 10,000</td>
-<td><span class="mstatus plan">structure representation, published</span></td>
+<td>structure-representation results in the paper</td>
 </tr>
 <tr>
 <td><a href="../examples/double-perovskite.html">Double perovskites A₂BB′X₆</a></td>
 <td>a family file, no dataset</td>
 <td>none</td>
 <td>10,800 compositions (halide variant)</td>
-<td><span class="mstatus part">design space and rules only</span></td>
+<td>design space and rules</td>
 </tr>
 <tr>
 <td><a href="#your-dataset">Your dataset</a></td>
 <td>one material family</td>
 <td>any scalar columns</td>
 <td>any</td>
-<td><span class="mstatus sup">supported</span></td>
+<td>upload in the Studio, or the command line</td>
 </tr>
 </tbody>
 </table>
@@ -2211,9 +2210,8 @@ with MACE and validated by DFT).</p>
 the <a href="https://babu09-meidnet.hf.space/studio/">live Studio</a>.</p>
 <p><strong>Suitable tasks:</strong> inverse design from a band gap and a stability target · property prediction from the
 structure · structure–property retrieval · measuring how well two modalities align.</p>
-<p><strong>MEIDNet implementation:</strong> ✓ pretrained model · ✓ configuration (<code>examples/perov5/meidnet.yaml</code>) ·
-✓ tutorial (<a href="../examples/perov5.html">the paper's experiment</a>, <a href="../start/colab.html">Colab</a>) ·
-✓ benchmark results.</p>
+<p><strong>Available here:</strong> pretrained model · configuration (<code>examples/perov5/meidnet.yaml</code>) ·
+tutorial (<a href="../examples/perov5.html">the paper's experiment</a>, <a href="../start/colab.html">Colab</a>) · benchmark results.</p>
 <p><a class="md-button md-button--primary" href="https://babu09-meidnet.hf.space/studio/">Open in the Studio</a></p>
 <p><a href="../benchmarks/perov5.html">Benchmark results</a> · <a href="../examples/perov5.html">The paper's experiment</a> ·
 <a href="https://github.com/txie-93/cdvae/tree/main/data/perov_5">Dataset source (CDVAE)</a> ·
@@ -2245,10 +2243,10 @@ or one <code>.cif</code> file per id) can be used directly: upload it in the <a
 or run <code>meidnet init</code> → <code>check</code> → <code>train</code> → <code>generate</code> (<a href="../use/your-data.html">bring your own dataset</a>).
 The structures must belong to one <a href="../reference/families.html">material family</a> (prototype + site groups), which is
 what the rules and the search need.</p>
-<h2 id="datasets-we-would-like-to-see-benchmarked">Datasets we would like to see benchmarked<a class="headerlink" href="#datasets-we-would-like-to-see-benchmarked" title="Permanent link">&para;</a></h2>
-<p>Battery conductors, MOFs, 2D materials and spinels are natural next families. If you run MEIDNet on one of them,
-<a href="../community/contribute.html">contribute the result</a>: the dataset gets its own benchmark page, and a verified row
-once the result has been reproduced here.</p>
+<h2 id="next-datasets">Next datasets<a class="headerlink" href="#next-datasets" title="Permanent link">&para;</a></h2>
+<p>Battery conductors, MOFs, 2D materials and spinels are natural next families. Results on any of them can be
+<a href="../community/contribute.html">contributed</a>: the dataset gets its own benchmark page, and a result reproduced here is
+marked as verified.</p>
 
 
 
diff --git a/site/index.html b/site/index.html
index cee816c1f5c94e3dddded0c53474cb6047d81b02..21665351dab892b381882565fa6ec9eb080d9018 100644
--- a/site/index.html
+++ b/site/index.html
@@ -540,10 +540,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#honest-scope" class="md-nav__link">
+  <a href="#scope" class="md-nav__link">
     <span class="md-ellipsis">
       
-        Honest scope
+        Scope
       
     </span>
   </a>
@@ -777,7 +777,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -976,7 +976,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1685,7 +1685,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2114,10 +2114,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#honest-scope" class="md-nav__link">
+  <a href="#scope" class="md-nav__link">
     <span class="md-ellipsis">
       
-        Honest scope
+        Scope
       
     </span>
   </a>
@@ -2160,7 +2160,8 @@
 <h1 id="meidnet-prism">MEIDNet Prism<a class="headerlink" href="#meidnet-prism" title="Permanent link">&para;</a></h1>
 <p><a href="https://babu09-meidnet.hf.space/">← MEIDNet Prism home</a> · the documentation of the MEIDNet framework and its platform.</p>
 <p><strong>Learn, build and benchmark multimodal AI for materials discovery.</strong> MEIDNet is the reference implementation:
-it designs crystalline materials from target properties, with your own data, your own rules and no code to edit.</p>
+it designs crystalline materials from target properties, using your own data and rules, configured in one YAML
+file or in the Studio.</p>
 <div class="highlight"><pre><span></span><code>   Learn ──► Architectures ──► Build ──► Datasets ──► Benchmarks ──► Community
    what is   which design      MEIDNet    data that    results with   share and
    multimodal fits your data   Studio     works        evidence       reproduce
@@ -2197,22 +2198,22 @@ that recommends one for your data and goal.</p>
 <li>
 <p><span class="twemoji lg middle"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path d="M10 16.5v-9l6 4.5M12 2A10 10 0 0 0 2 12a10 10 0 0 0 10 10 10 10 0 0 0 10-10A10 10 0 0 0 12 2"/></svg></span> <strong>Try MEIDNet</strong></p>
 <hr />
-<p>See what it does in 30 seconds: the <a href="https://babu09-meidnet.hf.space/studio/">live Studio</a> needs nothing installed;
-<code>meidnet demo</code> runs on any laptop.</p>
+<p>The <a href="https://babu09-meidnet.hf.space/studio/">live Studio</a> runs in the browser without installation;
+<code>meidnet demo</code> runs on a laptop CPU.</p>
 <p><a href="start/quickstart.html"><span class="twemoji"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path d="M13.22 19.03a.75.75 0 0 1 0-1.06L18.19 13H3.75a.75.75 0 0 1 0-1.5h14.44l-4.97-4.97a.749.749 0 0 1 .326-1.275.75.75 0 0 1 .734.215l6.25 6.25a.75.75 0 0 1 0 1.06l-6.25 6.25a.75.75 0 0 1-1.06 0"/></svg></span> 5-minute quickstart</a></p>
 </li>
 <li>
 <p><span class="twemoji lg middle"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path d="M12 3C8.59 3 5.69 4.07 4.54 5.57l5.25 5.25c.71.11 1.43.18 2.21.18 4.42 0 8-1.79 8-4s-3.58-4-8-4M3.92 7.08 2.5 8.5 5 11H0v2h5l-2.5 2.5 1.42 1.42L8.84 12M20 9c0 2.21-3.58 4-8 4-.66 0-1.3-.05-1.91-.13l-2.47 2.47c1.26.41 2.76.66 4.38.66 4.42 0 8-1.79 8-4m0 2c0 2.21-3.58 4-8 4-2.28 0-4.33-.5-5.79-1.25l-1.68 1.68C5.68 19.93 8.59 21 12 21c4.42 0 8-1.79 8-4"/></svg></span> <strong>Use MEIDNet on my data</strong></p>
 <hr />
-<p>A table with an id, properties and CIFs is all you need. Four commands take you from
-"is my data usable?" to candidates with explanations.</p>
+<p>A table with an id, property columns and CIFs. Four commands lead from a data check to
+candidates with reports.</p>
 <p><a href="use/your-data.html"><span class="twemoji"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path d="M13.22 19.03a.75.75 0 0 1 0-1.06L18.19 13H3.75a.75.75 0 0 1 0-1.5h14.44l-4.97-4.97a.749.749 0 0 1 .326-1.275.75.75 0 0 1 .734.215l6.25 6.25a.75.75 0 0 1 0 1.06l-6.25 6.25a.75.75 0 0 1-1.06 0"/></svg></span> Bring your own dataset</a></p>
 </li>
 <li>
 <p><span class="twemoji lg middle"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path d="M3 17v2h6v-2zM3 5v2h10V5zm10 16v-2h8v-2h-8v-2h-2v6zM7 9v2H3v2h4v2h2V9zm14 4v-2H11v2zm-6-4h2V7h4V5h-4V3h-2z"/></svg></span> <strong>Change what it does</strong></p>
 <hr />
-<p>Targets, elements, rules, family — all in one YAML file, or by moving sliders in
-MEIDNet Studio and exporting the file.</p>
+<p>Targets, elements, rules and family are set in one YAML file, or with the controls of
+MEIDNet Studio, which exports the same file.</p>
 <p><a href="recipes/change-targets.html"><span class="twemoji"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path d="M13.22 19.03a.75.75 0 0 1 0-1.06L18.19 13H3.75a.75.75 0 0 1 0-1.5h14.44l-4.97-4.97a.749.749 0 0 1 .326-1.275.75.75 0 0 1 .734.215l6.25 6.25a.75.75 0 0 1 0 1.06l-6.25 6.25a.75.75 0 0 1-1.06 0"/></svg></span> Recipes</a></p>
 </li>
 <li>
@@ -2229,9 +2230,9 @@ and where the limits are.</p>
    table    shared    prototype  hard      property   latent     CIFs + report
    + CIFs   latent    + sites    checks    goals      optimisation
 </code></pre></div>
-<p>Each block is a setting in <code>meidnet.yaml</code>, a node in <a href="use/studio.html">MEIDNet Studio</a> — where changing one
-block immediately shows its effect on the ones after it — and a section in the HTML report each step writes.</p>
-<h2 id="honest-scope">Honest scope<a class="headerlink" href="#honest-scope" title="Permanent link">&para;</a></h2>
+<p>Each block is a setting in <code>meidnet.yaml</code>, a node in <a href="use/studio.html">MEIDNet Studio</a>, where a change in one block
+updates the blocks after it, and a section in the HTML report that each step writes.</p>
+<h2 id="scope">Scope<a class="headerlink" href="#scope" title="Permanent link">&para;</a></h2>
 <ul>
 <li>Generation works for <strong>prototype families</strong>: a fixed arrangement of sites whose occupants and cell size are
   chosen (ABX₃ perovskites, A₂BB′X₆ double perovskites, and anything you describe the same way). MEIDNet does
@@ -2243,6 +2244,11 @@ block immediately shows its effect on the ones after it — and a section in the
 <p>A. Babu, R. Almeida Gouvêa, P. Vandergheynst, G.-M. Rignanese, <em>MEIDNet: Multimodal generative AI framework for
 inverse materials design</em>, npj Computational Materials (2026). <a href="research/paper.html">Citation and benchmarks</a> ·
 <a href="https://github.com/ABnano/MEIDNet/releases/tag/v1.0.0-paper">code as published (v1.0)</a>.</p>
+<p>Further reading: A. Babu, N. M. A. Krishnan, <em>Multimodal and cross-modal learning techniques</em>, APL Machine
+Learning <strong>4</strong>, 030901 (2026), <a href="https://doi.org/10.1063/5.0346744">doi:10.1063/5.0346744</a> · A. Babu,
+R. Almeida Gouvêa, G.-M. Rignanese, <em>Toward automated discovery with generative models multimodal learning and
+closed loop workflows in inverse materials design</em>, Cell Reports Physical Science <strong>7</strong>, 103561 (2026),
+<a href="https://doi.org/10.1016/j.xcrp.2026.103561">doi:10.1016/j.xcrp.2026.103561</a>.</p>
 
 
 
diff --git a/site/learn/architectures.html b/site/learn/architectures.html
index 0b2b1aa0375cc8a587a597b83e8a6ff49df9a30a..364d314d8f92f5246283e570dd03871aa5cb79b7 100644
--- a/site/learn/architectures.html
+++ b/site/learn/architectures.html
@@ -845,7 +845,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -1044,7 +1044,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1753,7 +1753,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2290,10 +2290,12 @@
 
 
 <h1 id="architecture-atlas">Architecture Atlas<a class="headerlink" href="#architecture-atlas" title="Permanent link">&para;</a></h1>
-<p>Five ways to build a model from several modalities, drawn the same way: inputs on the left, what the model produces
-on the right. Each card says when the design fits, what it costs, an example from materials science and whether
-MEIDNet implements it. At the end, <a href="#advisor">the advisor</a> recommends one for your data and your goal.</p>
-<p>New to the topic? Start with <a href="index.html">Learn multimodality</a>.</p>
+<p>Five ways to build a model from several modalities, drawn the same way: inputs on the left, outputs on the right.
+Each card gives the idea, when it fits, its strengths and weaknesses, an example from materials science and
+references. The <a href="#side-by-side">comparison table</a> shows which of these designs MEIDNet uses, and
+<a href="#advisor">the advisor</a> suggests one for a given dataset and task.</p>
+<p>New to the topic? Start with <a href="index.html">Learn multimodality</a>. Babu and Krishnan (2026) review multimodal and
+cross-modal learning techniques in more depth.</p>
 <div class="atlas">
 <div class="atlas-card">
 <div class="atlas-fig"><svg class="ad" viewBox="0 0 360 132" role="img" aria-label="Early fusion: features of two modalities are concatenated and given to one model">
@@ -2310,7 +2312,6 @@ MEIDNet implements it. At the end, <a href="#advisor">the advisor</a> recommends
 </svg></div>
 <div class="atlas-body">
 <h3 id="early-fusion">Early fusion<a class="headerlink" href="#early-fusion" title="Permanent link">&para;</a></h3>
-<p><span class="mstatus no">Not in MEIDNet</span></p>
 <p><strong>Idea.</strong> Join the modalities at the input: turn each into features, concatenate them into one vector, train one
 model on it.</p>
 <p><strong>When it fits.</strong> Every material has every modality; the features have a fixed length and similar scales; you want
@@ -2340,7 +2341,6 @@ pattern, concatenated and given to a gradient-boosted regressor.</p>
 </svg></div>
 <div class="atlas-body">
 <h3 id="late-fusion">Late fusion<a class="headerlink" href="#late-fusion" title="Permanent link">&para;</a></h3>
-<p><span class="mstatus no">Not in MEIDNet</span></p>
 <p><strong>Idea.</strong> One model per modality, each makes its own prediction; the predictions are combined by averaging, voting
 or a small model on top.</p>
 <p><strong>When it fits.</strong> Modalities are often missing, models for each already exist, or robustness matters more than
@@ -2373,7 +2373,6 @@ phase label; the final label averages their probabilities.</p>
 </svg></div>
 <div class="atlas-body">
 <h3 id="shared-latent">Shared latent space (joint and coordinated)<a class="headerlink" href="#shared-latent" title="Permanent link">&para;</a></h3>
-<p><span class="mstatus core">Core idea used by MEIDNet</span> <span class="mstatus sup">Supported</span></p>
 <p><strong>Idea.</strong> One encoder per modality maps into a common space. In a <em>coordinated</em> representation the latents stay
 separate but are trained to agree; in a <em>joint</em> representation they are merged into one vector that the decoders
 read. MEIDNet does both: the structure latent and the property latent are aligned by
@@ -2386,7 +2385,8 @@ encoder that suits its shape.
 <strong>Weaknesses.</strong> Needs paired data. The space is only as good as the alignment. Several losses must be balanced.</p>
 <p><strong>Materials example.</strong> MEIDNet: crystal structures and DFT properties of 18,928 cubic perovskites, searched for
 compositions with a target band gap and formation enthalpy (<a href="../examples/perov5.html">the paper's experiment</a>).</p>
-<p><em>References:</em> Ngiam <em>et al.</em> 2011; Baltrušaitis <em>et al.</em> 2019; Babu <em>et al.</em> 2026; Moro, Loh <em>et al.</em> 2025.</p>
+<p><em>References:</em> Ngiam <em>et al.</em> 2011; Baltrušaitis <em>et al.</em> 2019; Babu <em>et al.</em> 2026 (MEIDNet); Moro, Loh <em>et al.</em> 2025;
+Babu and Krishnan 2026.</p>
 </div>
 </div>
 <div class="atlas-card">
@@ -2406,7 +2406,6 @@ compositions with a target band gap and formation enthalpy (<a href="../examples
 </svg></div>
 <div class="atlas-body">
 <h3 id="cross-attention">Cross-attention and transformer fusion<a class="headerlink" href="#cross-attention" title="Permanent link">&para;</a></h3>
-<p><span class="mstatus no">Not in MEIDNet</span></p>
 <p><strong>Idea.</strong> Each modality becomes a sequence of tokens (atoms, segments of a spectrum, words). Attention lets every
 token of one modality look up the tokens of the other that matter to it, layer after layer.</p>
 <p><strong>When it fits.</strong> The modalities have parts that interact in detail and are not aligned in advance; the dataset is
@@ -2437,18 +2436,17 @@ which spectral features.</p>
 </svg></div>
 <div class="atlas-body">
 <h3 id="contrastive">Contrastive learning<a class="headerlink" href="#contrastive" title="Permanent link">&para;</a></h3>
-<p><span class="mstatus core">Used by MEIDNet</span> <span class="mstatus sup">Supported</span></p>
 <p><strong>Idea.</strong> A training objective rather than a wiring. In a batch of paired examples, each example's two embeddings
 must be more similar to each other than to any other example's. It is how the latents of a
 <a href="#shared-latent">shared space</a> are coordinated.</p>
 <p><strong>When it fits.</strong> You have pairs (a structure and its properties, a pattern and its structure) and want a space in
 which nearest neighbours are meaningful across modalities.</p>
-<p><strong>Strengths.</strong> Learns from the pairing alone; gives retrieval for free; scales to large datasets.
-<strong>Weaknesses.</strong> Needs batches with enough other examples; sensitive to the temperature; two genuinely similar
-materials in one batch are still pushed apart.</p>
+<p><strong>Strengths.</strong> Learns from the pairing alone; provides cross-modal retrieval; scales to large datasets.
+<strong>Weaknesses.</strong> Needs large enough batches; sensitive to the temperature; two similar materials in the same batch
+are treated as negatives and pushed apart.</p>
 <p><strong>In MEIDNet.</strong> Symmetric InfoNCE with temperature 0.01 and weight 5, switched on gradually over the warm-up
 (the paper's curriculum). <a href="index.html#contrastive-learning">The formula and a playground</a>.</p>
-<p><em>References:</em> van den Oord <em>et al.</em> 2018 (InfoNCE); Radford <em>et al.</em> 2021 (CLIP).</p>
+<p><em>References:</em> van den Oord <em>et al.</em> 2018 (InfoNCE); Radford <em>et al.</em> 2021 (CLIP); Babu and Krishnan 2026.</p>
 </div>
 </div>
 </div>
@@ -2506,21 +2504,21 @@ materials in one batch are still pushed apart.</p>
 <td>medium to large</td>
 </tr>
 <tr>
-<td>in MEIDNet</td>
-<td>no</td>
-<td>no</td>
-<td><strong>yes, the core</strong></td>
-<td>no</td>
-<td><strong>yes</strong></td>
+<td>used in MEIDNet</td>
+<td>–</td>
+<td>–</td>
+<td>yes: structure and property encoders, averaged latent</td>
+<td>–</td>
+<td>yes: symmetric InfoNCE</td>
 </tr>
 </tbody>
 </table>
 <p><strong>Beyond these five.</strong> Conditional generative models such as diffusion (MatterGen) or variational autoencoders
-(CDVAE) also translate from properties to structures, and they generate free atomic arrangements, which MEIDNet does
-not. MEIDNet instead decodes compositions onto a prototype family and checks them against chemical rules.</p>
+(CDVAE) also translate from properties to structures, and they generate free atomic arrangements. MEIDNet instead
+decodes compositions onto a prototype family and checks them against chemical rules. Babu, Gouvêa and Rignanese
+(2026) review generative models, multimodal learning and closed-loop workflows for inverse materials design.</p>
 <h2 id="advisor">Which architecture should I use?<a class="headerlink" href="#advisor" title="Permanent link">&para;</a></h2>
-<p>Tick the data you have and choose what you want to do. The recommendation says why, what it costs and whether
-MEIDNet can do it today.</p>
+<p>Select the data you have and the task. The recommendation explains the choice and says where it can be run.</p>
 <div class="advisor" id="arch-advisor" aria-live="polite"></div>
 
 <noscript>The advisor needs JavaScript. In short: structures with scalar properties, for design or retrieval →
@@ -2562,6 +2560,11 @@ over <code>training.contrastive_warmup_epochs</code>, so that the decoders learn
   (2022). <a href="https://arxiv.org/abs/2110.06197">arXiv:2110.06197</a></li>
 <li>V. Moro, C. Loh <em>et al.</em>, "Multimodal foundation models for material property prediction and discovery",
   <em>Newton</em> (2025). <a href="https://doi.org/10.1016/j.newton.2025.100016">doi:10.1016/j.newton.2025.100016</a></li>
+<li>A. Babu, N. M. A. Krishnan, "Multimodal and cross-modal learning techniques", <em>APL Mach. Learn.</em> <strong>4</strong>, 030901
+  (2026). <a href="https://doi.org/10.1063/5.0346744">doi:10.1063/5.0346744</a></li>
+<li>A. Babu, R. Almeida Gouvêa, G.-M. Rignanese, "Toward automated discovery with generative models multimodal
+  learning and closed loop workflows in inverse materials design", <em>Cell Rep. Phys. Sci.</em> <strong>7</strong>, 103561 (2026).
+  <a href="https://doi.org/10.1016/j.xcrp.2026.103561">doi:10.1016/j.xcrp.2026.103561</a></li>
 <li>A. Babu, R. Almeida Gouvêa, P. Vandergheynst, G.-M. Rignanese, "MEIDNet: Multimodal generative AI framework for
   inverse materials design", <em>npj Comput. Mater.</em> (2026). <a href="https://doi.org/10.1038/s41524-026-02153-3">doi:10.1038/s41524-026-02153-3</a></li>
 </ul>
diff --git a/site/learn/index.html b/site/learn/index.html
index b6f2dba38d5f37e2f30f643c9e5cdc9a5c1dab66..137261d08861255c79433c9756ac44e3751b4ff8 100644
--- a/site/learn/index.html
+++ b/site/learn/index.html
@@ -828,7 +828,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -1027,7 +1027,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1736,7 +1736,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2254,10 +2254,10 @@
 <p>A material can be described in many ways: by its atoms and their positions, by a formula, by measured or computed
 properties, by a diffraction pattern, a spectrum, a micrograph or a paragraph in a paper. Each way of describing it
 is a <strong>modality</strong>. Multimodal learning trains one model on several modalities of the same materials, so that it can
-connect them: read a structure and predict a property, or — the point of MEIDNet — start from the properties you
-want and propose a structure.</p>
-<p>This short course takes about ten minutes. It explains the ideas behind MEIDNet and the alternatives to it, and every
-item says what works in MEIDNet today.</p>
+connect them: read a structure and predict a property, or start from the properties you want and propose a
+structure, which is what MEIDNet does.</p>
+<p>This short course takes about ten minutes. It covers the main ideas of multimodal learning for materials and, for
+each, how MEIDNet implements it. For a full review, see Babu and Krishnan (2026).</p>
 <div class="learn-steps">
 <ol>
 <li><a href="#what-is-a-modality">What is a modality?</a></li>
@@ -2331,8 +2331,8 @@ together can translate between them in either direction.</p>
 <p><strong>What it needs:</strong> <em>paired</em> data, where several modalities describe the same material. In Perov-5, every one of the
 18,928 structures comes with its DFT band gap and formation enthalpy, which makes it a two-modality dataset.</p>
 <h2 id="the-multimodality-map">2. The multimodality map<a class="headerlink" href="#the-multimodality-map" title="Permanent link">&para;</a></h2>
-<p>Click a modality to see what it is, which networks read it, how it is combined with others and its status in
-MEIDNet. Click the centre for the idea that ties them together.</p>
+<p>Select a modality to see what it contains, which networks encode it, how it is combined with other modalities and
+how MEIDNet handles it. The centre describes the shared representation.</p>
 <div class="mm" id="mm-map" aria-label="Interactive map of materials modalities"></div>
 
 <details class="note">
@@ -2341,54 +2341,54 @@ MEIDNet. Click the centre for the idea that ties them together.</p>
 <thead>
 <tr>
 <th>modality</th>
-<th>status in MEIDNet</th>
+<th>MEIDNet</th>
 <th>how</th>
 </tr>
 </thead>
 <tbody>
 <tr>
 <td>Crystal structure</td>
-<td><strong>Supported</strong></td>
+<td>available</td>
 <td>an E(n)-equivariant graph network (EGNN) encodes up to <code>max_sites</code> atoms (default 20)</td>
 </tr>
 <tr>
 <td>Scalar properties</td>
-<td><strong>Supported</strong></td>
+<td>available</td>
 <td>all numeric columns form one vector, read by one MLP; every column becomes a target</td>
 </tr>
 <tr>
 <td>Composition</td>
-<td><strong>Inside the structure</strong></td>
+<td>via the structure</td>
 <td>the element on each site is part of the structure; formula-only data needs a composition model</td>
 </tr>
 <tr>
 <td>Processing conditions</td>
-<td><strong>Partly</strong></td>
+<td>partial</td>
 <td>numeric conditions can be extra scalar columns, encoded together with the properties</td>
 </tr>
 <tr>
 <td>Diffraction (XRD)</td>
-<td><strong>Planned</strong></td>
+<td>planned</td>
 <td>a vector-modality encoder for binned patterns, first item of the <a href="../understand/limits.html">roadmap</a></td>
 </tr>
 <tr>
 <td>Spectra and DOS</td>
-<td><strong>Planned</strong></td>
+<td>planned</td>
 <td>as vector modalities</td>
 </tr>
 <tr>
 <td>Microscopy images</td>
-<td><strong>Planned</strong></td>
+<td>planned</td>
 <td>an image encoder into the shared space</td>
 </tr>
 <tr>
 <td>Text</td>
-<td><strong>Planned</strong></td>
+<td>planned</td>
 <td>a text encoder into the shared space</td>
 </tr>
 </tbody>
 </table>
-<p>The same statuses, generated from the code, are on the <a href="../explore/capabilities.html">capabilities page</a>.</p>
+<p>The <a href="../explore/capabilities.html">capabilities page</a> lists the same information, generated from the code.</p>
 </details>
 <h2 id="five-challenges">3. Five challenges of multimodal learning<a class="headerlink" href="#five-challenges" title="Permanent link">&para;</a></h2>
 <p>Baltrušaitis, Ahuja and Morency organise multimodal machine learning around five challenges. Every multimodal model
@@ -2399,7 +2399,7 @@ answers each of them in some way, even if only by leaving it out.</p>
 <th>challenge</th>
 <th>the question</th>
 <th>a materials example</th>
-<th>the answer in MEIDNet</th>
+<th>in MEIDNet</th>
 </tr>
 </thead>
 <tbody>
@@ -2436,7 +2436,8 @@ answers each of them in some way, even if only by leaving it out.</p>
 </tbody>
 </table>
 <p><em>Further reading:</em> Baltrušaitis <em>et al.</em> 2019 (the taxonomy) · Guo <em>et al.</em> 2019 (deep multimodal
-representation learning) · Ngiam <em>et al.</em> 2011 (shared representations between modalities).</p>
+representation learning) · Ngiam <em>et al.</em> 2011 (shared representations between modalities) · Babu and Krishnan
+2026 (multimodal and cross-modal learning techniques).</p>
 <h2 id="contrastive-learning">4. Contrastive learning: how two modalities are aligned<a class="headerlink" href="#contrastive-learning" title="Permanent link">&para;</a></h2>
 <p>Take a batch of <span class="arithmatex">\(N\)</span> materials. Each gives two vectors of length one: <span class="arithmatex">\(z_c\)</span> from its structure and <span class="arithmatex">\(z_p\)</span> from its
 properties. The dot product of two such vectors is their cosine similarity. Lay out all <span class="arithmatex">\(N \times N\)</span> similarities
@@ -2490,7 +2491,7 @@ latents are pulled together. After training, the <a href="../use/reports.html">t
 <em>retrieval</em>: how often the nearest property vector to a structure is its own (top 1) or among the five nearest
 (top 5). The Studio's Model block plots the mean cosine between matched pairs during training.</p>
 <p>The same loss pairs pictures with their captions in CLIP; MEIDNet pairs crystals with their properties.</p>
-<p><em>Further reading:</em> van den Oord <em>et al.</em> 2018 (InfoNCE) · Radford <em>et al.</em> 2021 (CLIP) ·
+<p><em>Further reading:</em> van den Oord <em>et al.</em> 2018 (InfoNCE) · Radford <em>et al.</em> 2021 (CLIP) · Babu and Krishnan 2026 ·
 <a href="architectures.html#contrastive">Contrastive learning in the Architecture Atlas</a>.</p>
 <h2 id="inverse-design">5. From a shared space to inverse design<a class="headerlink" href="#inverse-design" title="Permanent link">&para;</a></h2>
 <p>Once both modalities live in one space, a point in it can be read in both directions. MEIDNet makes the reverse
@@ -2509,8 +2510,9 @@ alone</strong>. A vector of wanted properties can then be encoded and decoded in
 <p>The full objective and every step are in <a href="../understand/how-it-works.html">How MEIDNet works</a>. Two limits matter:
 MEIDNet does not invent new atomic arrangements, and its predicted properties are estimates to confirm with DFT or
 experiment (<a href="../understand/limits.html">scope and roadmap</a>).</p>
-<p><em>Further reading:</em> the MEIDNet paper (Babu <em>et al.</em> 2026) · Moro, Loh <em>et al.</em> 2025 (multimodal foundation models
-for materials).</p>
+<p><em>Further reading:</em> the MEIDNet paper (Babu <em>et al.</em> 2026) · Babu, Gouvêa and Rignanese 2026 (generative models,
+multimodal learning and closed-loop workflows in inverse materials design) · Moro, Loh <em>et al.</em> 2025 (multimodal
+foundation models for materials).</p>
 <h2 id="where-next">6. Where to go next<a class="headerlink" href="#where-next" title="Permanent link">&para;</a></h2>
 <div class="grid cards">
 <ul>
@@ -2556,6 +2558,11 @@ with an advisor that recommends one for your data.</p>
   <a href="https://arxiv.org/abs/2102.09844">arXiv:2102.09844</a></li>
 <li>V. Moro, C. Loh <em>et al.</em>, "Multimodal foundation models for material property prediction and discovery",
   <em>Newton</em> (2025). <a href="https://doi.org/10.1016/j.newton.2025.100016">doi:10.1016/j.newton.2025.100016</a></li>
+<li>A. Babu, N. M. A. Krishnan, "Multimodal and cross-modal learning techniques", <em>APL Mach. Learn.</em> <strong>4</strong>, 030901
+  (2026). <a href="https://doi.org/10.1063/5.0346744">doi:10.1063/5.0346744</a></li>
+<li>A. Babu, R. Almeida Gouvêa, G.-M. Rignanese, "Toward automated discovery with generative models multimodal
+  learning and closed loop workflows in inverse materials design", <em>Cell Rep. Phys. Sci.</em> <strong>7</strong>, 103561 (2026).
+  <a href="https://doi.org/10.1016/j.xcrp.2026.103561">doi:10.1016/j.xcrp.2026.103561</a></li>
 <li>A. Babu, R. Almeida Gouvêa, P. Vandergheynst, G.-M. Rignanese, "MEIDNet: Multimodal generative AI framework for
   inverse materials design", <em>npj Comput. Mater.</em> (2026). <a href="https://doi.org/10.1038/s41524-026-02153-3">doi:10.1038/s41524-026-02153-3</a></li>
 </ul>
diff --git a/site/learn/recipes.html b/site/learn/recipes.html
index 29520a271bf9df73e02c2717e1f7b4649c7d57bb..82fc5a15265eb44c9061ab42211a46829f330bbf 100644
--- a/site/learn/recipes.html
+++ b/site/learn/recipes.html
@@ -839,7 +839,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -1038,7 +1038,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1747,7 +1747,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2273,9 +2273,9 @@
 
 
 <h1 id="recipes-by-problem">Recipes by problem<a class="headerlink" href="#recipes-by-problem" title="Permanent link">&para;</a></h1>
-<p>Start from a research question. Each recipe goes <strong>problem → architecture → why → data → configuration → tutorial
-→ references</strong>, and says plainly whether MEIDNet does it today. Where it does not, the recipe points to the kind of
-model that does.</p>
+<p>Start from a research question. Each recipe goes <strong>problem → architecture → why → data → implementation →
+tutorial → references</strong>. The implementation is MEIDNet where it applies; otherwise the recipe names established
+alternatives.</p>
 <p>For step-by-step changes to a MEIDNet configuration (another target, another family, a new rule), see the
 <a href="../recipes/change-targets.html">how-to guides</a>.</p>
 <table>
@@ -2283,49 +2283,48 @@ model that does.</p>
 <tr>
 <th>problem</th>
 <th>architecture</th>
-<th>MEIDNet today</th>
+<th>implementation</th>
 </tr>
 </thead>
 <tbody>
 <tr>
 <td><a href="#target-band-gap">Perovskites with a target band gap that are likely to be stable</a></td>
 <td>shared latent + contrastive</td>
-<td><span class="mstatus sup">Supported</span></td>
+<td>MEIDNet</td>
 </tr>
 <tr>
 <td><a href="#own-family">Candidates from your own DFT results for one family</a></td>
 <td>shared latent + contrastive</td>
-<td><span class="mstatus sup">Supported</span></td>
+<td>MEIDNet</td>
 </tr>
 <tr>
 <td><a href="#screen-family">Screen a whole family for a target before searching</a></td>
 <td>structure encoder of the shared space</td>
-<td><span class="mstatus sup">Supported</span></td>
+<td>MEIDNet</td>
 </tr>
 <tr>
 <td><a href="#xrd-structure">Identify the crystal structure from an XRD pattern</a></td>
 <td>1D convolutional network; contrastive pattern ↔ structure</td>
-<td><span class="mstatus plan">Planned</span></td>
+<td>1D CNN classifiers; pattern encoder planned for MEIDNet</td>
 </tr>
 <tr>
 <td><a href="#formula-only">Predict a property when only the formula is known</a></td>
 <td>composition network</td>
-<td><span class="mstatus no">Use another tool</span></td>
+<td>Roost and similar composition models</td>
 </tr>
 <tr>
 <td><a href="#synthesis-text">Connect synthesis text with structures</a></td>
 <td>text encoder + contrastive or cross-attention</td>
-<td><span class="mstatus plan">Planned</span></td>
+<td>text encoder planned for MEIDNet</td>
 </tr>
 <tr>
 <td><a href="#new-arrangements">Design new atomic arrangements</a></td>
 <td>conditional generative model over full geometry</td>
-<td><span class="mstatus no">Use another tool</span></td>
+<td>MatterGen, CDVAE</td>
 </tr>
 </tbody>
 </table>
 <h2 id="target-band-gap">Perovskites with a target band gap that are likely to be stable<a class="headerlink" href="#target-band-gap" title="Permanent link">&para;</a></h2>
-<p><span class="mstatus sup">Supported in MEIDNet</span></p>
 <table>
 <thead>
 <tr>
@@ -2351,12 +2350,16 @@ model that does.</p>
 <td>Perov-5: 18,928 structures with DFT band gap and formation enthalpy (<a href="../explore/datasets.html">dataset card</a>). The published checkpoint is trained on it.</td>
 </tr>
 <tr>
+<td><strong>Implementation</strong></td>
+<td>MEIDNet: <code>meidnet demo</code>, <code>meidnet generate</code> or the Studio's Targets and Search blocks.</td>
+</tr>
+<tr>
 <td><strong>Tutorial</strong></td>
 <td><a href="../examples/perov5.html">The paper's experiment</a> · in the browser: <a href="https://babu09-meidnet.hf.space/studio/?panel=targets">Studio, Targets block</a> · <a href="../recipes/screen.html">screen the candidates with MACE</a></td>
 </tr>
 <tr>
 <td><strong>References</strong></td>
-<td>Babu <em>et al.</em> 2026 (MEIDNet); Castelli <em>et al.</em> 2012 (the dataset).</td>
+<td>Babu <em>et al.</em> 2026 (MEIDNet); Castelli <em>et al.</em> 2012 (the dataset); Babu, Gouvêa and Rignanese 2026 (review of inverse design).</td>
 </tr>
 </tbody>
 </table>
@@ -2374,7 +2377,6 @@ model that does.</p>
 <span class="w">    </span><span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">{</span><span class="nt">dir_gap</span><span class="p">:</span><span class="w"> </span><span class="nv">1.5</span><span class="p p-Indicator">,</span><span class="nt"> heat_all</span><span class="p">:</span><span class="w"> </span><span class="nv">-0.10</span><span class="p p-Indicator">}</span>
 </code></pre></div>
 <h2 id="own-family">Candidates from your own DFT results for one family<a class="headerlink" href="#own-family" title="Permanent link">&para;</a></h2>
-<p><span class="mstatus sup">Supported in MEIDNet</span></p>
 <table>
 <thead>
 <tr>
@@ -2400,12 +2402,16 @@ model that does.</p>
 <td>A table with an id, numeric property columns and one CIF per row. Public sources by application are on <a href="../explore/databases.html">Databases by application</a>.</td>
 </tr>
 <tr>
+<td><strong>Implementation</strong></td>
+<td>MEIDNet: <code>meidnet init</code>, <code>check</code>, <code>train</code>, <code>generate</code>, or the Studio's Data and Model blocks.</td>
+</tr>
+<tr>
 <td><strong>Tutorial</strong></td>
 <td><a href="../use/your-data.html">Bring your own dataset</a> · in the browser: <a href="https://babu09-meidnet.hf.space/studio/?panel=data">Studio, Data block</a> · <a href="../recipes/change-family.html">change the material family</a></td>
 </tr>
 <tr>
 <td><strong>References</strong></td>
-<td>Babu <em>et al.</em> 2026.</td>
+<td>Babu <em>et al.</em> 2026 (MEIDNet).</td>
 </tr>
 </tbody>
 </table>
@@ -2427,7 +2433,6 @@ model that does.</p>
 </code></pre></div>
 <p>Then <code>meidnet check</code> → <code>meidnet train</code> → <code>meidnet generate</code>; each step writes a report.</p>
 <h2 id="screen-family">Screen a whole family for a target before searching<a class="headerlink" href="#screen-family" title="Permanent link">&para;</a></h2>
-<p><span class="mstatus sup">Supported in MEIDNet</span></p>
 <table>
 <thead>
 <tr>
@@ -2450,7 +2455,11 @@ model that does.</p>
 </tr>
 <tr>
 <td><strong>Data</strong></td>
-<td>A trained model and a family; nothing else.</td>
+<td>A trained model and a family.</td>
+</tr>
+<tr>
+<td><strong>Implementation</strong></td>
+<td>MEIDNet: <code>meidnet space</code> or the Studio's design-space view.</td>
 </tr>
 <tr>
 <td><strong>Tutorial</strong></td>
@@ -2458,14 +2467,13 @@ model that does.</p>
 </tr>
 <tr>
 <td><strong>References</strong></td>
-<td>Babu <em>et al.</em> 2026.</td>
+<td>Babu <em>et al.</em> 2026 (MEIDNet).</td>
 </tr>
 </tbody>
 </table>
 <div class="highlight"><pre><span></span><code>meidnet<span class="w"> </span>space<span class="w"> </span>examples/perov5/meidnet.yaml<span class="w"> </span>--model<span class="w"> </span>checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth<span class="w"> </span>-o<span class="w"> </span>space.csv
 </code></pre></div>
 <h2 id="xrd-structure">Identify the crystal structure from an XRD pattern<a class="headerlink" href="#xrd-structure" title="Permanent link">&para;</a></h2>
-<p><span class="mstatus plan">Planned in MEIDNet</span></p>
 <table>
 <thead>
 <tr>
@@ -2480,7 +2488,7 @@ model that does.</p>
 </tr>
 <tr>
 <td><strong>Architecture</strong></td>
-<td>Today: a 1D convolutional classifier on the pattern (one modality). Multimodal: a pattern encoder aligned with a structure encoder by <a href="architectures.html#contrastive">contrastive learning</a>, so a pattern retrieves the structures it matches.</td>
+<td>A 1D convolutional classifier on the pattern (one modality). Multimodal: a pattern encoder aligned with a structure encoder by <a href="architectures.html#contrastive">contrastive learning</a>, so a pattern retrieves the structures it matches.</td>
 </tr>
 <tr>
 <td><strong>Why</strong></td>
@@ -2491,17 +2499,16 @@ model that does.</p>
 <td>Simulate patterns from computed or experimental structures (<a href="../explore/databases.html#semiconductor-physics">semiconductor physics</a>, <a href="../explore/databases.html#all-databases">all databases</a>), for example with pymatgen's XRD calculator; validate on measured patterns.</td>
 </tr>
 <tr>
-<td><strong>In MEIDNet</strong></td>
-<td>A vector-modality encoder for binned XRD is the first item of the <a href="../understand/limits.html">roadmap</a>. It is not implemented.</td>
+<td><strong>Implementation</strong></td>
+<td>1D convolutional classifiers such as Park <em>et al.</em> (2017). In MEIDNet, an encoder for binned XRD patterns is the first item of the <a href="../understand/limits.html">roadmap</a> and is not yet available.</td>
 </tr>
 <tr>
 <td><strong>References</strong></td>
-<td>Park <em>et al.</em> 2017; Baltrušaitis <em>et al.</em> 2019.</td>
+<td>Park <em>et al.</em> 2017; Baltrušaitis <em>et al.</em> 2019; Babu and Krishnan 2026.</td>
 </tr>
 </tbody>
 </table>
 <h2 id="formula-only">Predict a property when only the formula is known<a class="headerlink" href="#formula-only" title="Permanent link">&para;</a></h2>
-<p><span class="mstatus no">Use another tool</span></p>
 <table>
 <thead>
 <tr>
@@ -2520,20 +2527,23 @@ model that does.</p>
 </tr>
 <tr>
 <td><strong>Why</strong></td>
-<td>MEIDNet needs a structure for every row. If your formulas all sit on one prototype (for example ABX₃), you can build those structures on the prototype and use MEIDNet; otherwise a composition model is the right tool.</td>
+<td>MEIDNet needs a structure for every row. If your formulas all sit on one prototype (for example ABX₃), you can build those structures on the prototype and use MEIDNet; otherwise a composition model is the appropriate choice.</td>
 </tr>
 <tr>
 <td><strong>Data</strong></td>
 <td>Experimental band gaps and Matbench tasks are listed on <a href="../explore/databases.html#semiconductor-physics">Databases by application</a>.</td>
 </tr>
 <tr>
+<td><strong>Implementation</strong></td>
+<td>Roost or another composition model; outside MEIDNet, which requires a crystal structure per material.</td>
+</tr>
+<tr>
 <td><strong>References</strong></td>
 <td>Goodall and Lee 2020 (Roost); Dunn <em>et al.</em> 2020 (Matbench).</td>
 </tr>
 </tbody>
 </table>
 <h2 id="synthesis-text">Connect synthesis text with structures<a class="headerlink" href="#synthesis-text" title="Permanent link">&para;</a></h2>
-<p><span class="mstatus plan">Planned in MEIDNet</span></p>
 <table>
 <thead>
 <tr>
@@ -2559,17 +2569,16 @@ model that does.</p>
 <td>Text-mined synthesis recipes (Kononova <em>et al.</em> 2019) paired with structures from a structure database.</td>
 </tr>
 <tr>
-<td><strong>In MEIDNet</strong></td>
-<td>A text encoder into the shared space is planned, not implemented.</td>
+<td><strong>Implementation</strong></td>
+<td>A text encoder that maps into MEIDNet's shared space is planned and not yet available.</td>
 </tr>
 <tr>
 <td><strong>References</strong></td>
-<td>Kononova <em>et al.</em> 2019; Moro, Loh <em>et al.</em> 2025.</td>
+<td>Kononova <em>et al.</em> 2019; Moro, Loh <em>et al.</em> 2025; Babu and Krishnan 2026.</td>
 </tr>
 </tbody>
 </table>
 <h2 id="new-arrangements">Design new atomic arrangements<a class="headerlink" href="#new-arrangements" title="Permanent link">&para;</a></h2>
-<p><span class="mstatus no">Use another tool</span></p>
 <table>
 <thead>
 <tr>
@@ -2595,8 +2604,12 @@ model that does.</p>
 <td>Large structure databases; the generative benchmarks are listed under <a href="../explore/databases.html#generative-model-benchmarks">generative-model benchmarks</a>.</td>
 </tr>
 <tr>
+<td><strong>Implementation</strong></td>
+<td>MatterGen, CDVAE and related generative models.</td>
+</tr>
+<tr>
 <td><strong>References</strong></td>
-<td>Zeni <em>et al.</em> 2025 (MatterGen); Xie <em>et al.</em> 2022 (CDVAE).</td>
+<td>Zeni <em>et al.</em> 2025 (MatterGen); Xie <em>et al.</em> 2022 (CDVAE); Babu, Gouvêa and Rignanese 2026 (review).</td>
 </tr>
 </tbody>
 </table>
@@ -2614,6 +2627,11 @@ model that does.</p>
   test set and Automatminer reference algorithm", <em>npj Comput. Mater.</em> <strong>6</strong>, 138 (2020). <a href="https://doi.org/10.1038/s41524-020-00406-3">doi:10.1038/s41524-020-00406-3</a></li>
 <li>O. Kononova <em>et al.</em>, "Text-mined dataset of inorganic materials synthesis recipes", <em>Sci. Data</em> <strong>6</strong>, 203 (2019).
   <a href="https://doi.org/10.1038/s41597-019-0224-1">doi:10.1038/s41597-019-0224-1</a></li>
+<li>A. Babu, N. M. A. Krishnan, "Multimodal and cross-modal learning techniques", <em>APL Mach. Learn.</em> <strong>4</strong>, 030901
+  (2026). <a href="https://doi.org/10.1063/5.0346744">doi:10.1063/5.0346744</a></li>
+<li>A. Babu, R. Almeida Gouvêa, G.-M. Rignanese, "Toward automated discovery with generative models multimodal
+  learning and closed loop workflows in inverse materials design", <em>Cell Rep. Phys. Sci.</em> <strong>7</strong>, 103561 (2026).
+  <a href="https://doi.org/10.1016/j.xcrp.2026.103561">doi:10.1016/j.xcrp.2026.103561</a></li>
 <li>C. Zeni <em>et al.</em>, "A generative model for inorganic materials design", <em>Nature</em> <strong>639</strong>, 624–632 (2025).
   <a href="https://doi.org/10.1038/s41586-025-08628-5">doi:10.1038/s41586-025-08628-5</a></li>
 <li>T. Xie <em>et al.</em>, "Crystal diffusion variational autoencoder for periodic material generation", <em>ICLR</em> (2022).
diff --git a/site/recipes/add-constraint.html b/site/recipes/add-constraint.html
index b37894976eae04588cc355b9ffaf74f286f35a21..107ec9a20ffe6d417ee8613e97b63866d6a08b71 100644
--- a/site/recipes/add-constraint.html
+++ b/site/recipes/add-constraint.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1703,7 +1703,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/recipes/add-property.html b/site/recipes/add-property.html
index 32eac4a15b5a9f605bb39989e1978a6bfc9a1b36..11eebe15eb752fe504d385d7a932f88a5547c1da 100644
--- a/site/recipes/add-property.html
+++ b/site/recipes/add-property.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1622,7 +1622,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/recipes/change-family.html b/site/recipes/change-family.html
index 20284dffe6c093d81dbdefdaa22effb772db25e4..3f62f39b63ca19742ba88296f095b1d61aa10fc7 100644
--- a/site/recipes/change-family.html
+++ b/site/recipes/change-family.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1703,7 +1703,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/recipes/change-targets.html b/site/recipes/change-targets.html
index 198e9785d81884b10de2b1cbd5f40199807cf99a..0f2e5cdca3ad17c8dfbc67210982a640796b0591 100644
--- a/site/recipes/change-targets.html
+++ b/site/recipes/change-targets.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1703,7 +1703,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/recipes/elements.html b/site/recipes/elements.html
index ab797d1427f724f14fa3ea4eecd08eebe2c1f1f9..f37ab8fd9be3532e557575c790210a58371eaf3a 100644
--- a/site/recipes/elements.html
+++ b/site/recipes/elements.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1622,7 +1622,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/recipes/screen.html b/site/recipes/screen.html
index 9937067dc341e8638c21f4916c7fdf917d40b6a9..c29484bcfa95bd0eecffbe510945c000507aec55 100644
--- a/site/recipes/screen.html
+++ b/site/recipes/screen.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1622,7 +1622,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/reference/api.html b/site/reference/api.html
index a68077c69080faa5194487305df434e642d7ac56..acfc18a16ffb85133ca13fa4eb4ef0abe1d2da40 100644
--- a/site/reference/api.html
+++ b/site/reference/api.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1607,7 +1607,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/reference/cli.html b/site/reference/cli.html
index 1b6620cefc20160ab870a7384d0fcb5ae6c2d49a..0159f81df6bd211bed9789653fef84606fde61c1 100644
--- a/site/reference/cli.html
+++ b/site/reference/cli.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1607,7 +1607,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/reference/config.html b/site/reference/config.html
index d185a0081349612495006fc44daebb6e4a114b0c..8c72098fe86002e0771c95d966a4ccb253c8afcf 100644
--- a/site/reference/config.html
+++ b/site/reference/config.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1607,7 +1607,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/reference/families.html b/site/reference/families.html
index 90cd2ac7b74d792db060fc8926c6ac1e932e06e4..a4008a860e7a9deb082eec8eca312ade73ab04df 100644
--- a/site/reference/families.html
+++ b/site/reference/families.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1607,7 +1607,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/reference/rules.html b/site/reference/rules.html
index cc78ada011e66c5d7a02b1e0bf8fc0c5ffe9f4bc..7d9ff45d71770f266e479fa83bbe4289db5f69bf 100644
--- a/site/reference/rules.html
+++ b/site/reference/rules.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1607,7 +1607,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/research/paper.html b/site/research/paper.html
index f6e2b05919113598834d0ddbc4b6fe398f2e73fe..516263f75a41d3387c813dc5cd9714a1b299be02 100644
--- a/site/research/paper.html
+++ b/site/research/paper.html
@@ -697,7 +697,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -896,7 +896,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1605,7 +1605,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2161,6 +2161,15 @@ Anand Babu, Rogério Almeida Gouvêa, Pierre Vandergheynst, Gian-Marco Rignanese
 <li>Generated structures: <code>examples/perov5/paper_results/</code>.</li>
 </ul>
 <h2 id="related">Related<a class="headerlink" href="#related" title="Permanent link">&para;</a></h2>
+<p>Reviews by the authors:</p>
+<ul>
+<li>A. Babu, N. M. A. Krishnan, <em>Multimodal and cross-modal learning techniques</em>, APL Machine Learning <strong>4</strong>, 030901
+  (2026). <a href="https://doi.org/10.1063/5.0346744">doi:10.1063/5.0346744</a></li>
+<li>A. Babu, R. Almeida Gouvêa, G.-M. Rignanese, <em>Toward automated discovery with generative models multimodal
+  learning and closed loop workflows in inverse materials design</em>, Cell Reports Physical Science <strong>7</strong>, 103561
+  (2026). <a href="https://doi.org/10.1016/j.xcrp.2026.103561">doi:10.1016/j.xcrp.2026.103561</a></li>
+</ul>
+<p>Methods MEIDNet builds on:</p>
 <ul>
 <li>Satorras, Hoogeboom, Welling, <em>E(n) equivariant graph neural networks</em>, ICML 2021 (the encoder).</li>
 <li>Radford et al., <em>CLIP</em>, ICML 2021 (the alignment objective).</li>
diff --git a/site/search/search_index.json b/site/search/search_index.json
index 44f97978b54728e3bc963908a8a41f94a9e54ea8..771361ffb35d5dd91ea0f63d0fdaa0d8ab766d76 100644
--- a/site/search/search_index.json
+++ b/site/search/search_index.json
@@ -1 +1 @@
-{"config":{"lang":["en"],"separator":"[\\s\\-]+","pipeline":["stopWordFilter"],"fields":{"title":{"boost":1000.0},"text":{"boost":1.0},"tags":{"boost":1000000.0}}},"docs":[{"location":"index.html","title":"Home","text":""},{"location":"index.html#meidnet-prism","title":"MEIDNet Prism","text":"<p>\u2190 MEIDNet Prism home \u00b7 the documentation of the MEIDNet framework and its platform.</p> <p>Learn, build and benchmark multimodal AI for materials discovery. MEIDNet is the reference implementation: it designs crystalline materials from target properties, with your own data, your own rules and no code to edit.</p> <pre><code>   Learn \u2500\u2500\u25ba Architectures \u2500\u2500\u25ba Build \u2500\u2500\u25ba Datasets \u2500\u2500\u25ba Benchmarks \u2500\u2500\u25ba Community\n   what is   which design      MEIDNet    data that    results with   share and\n   multimodal fits your data   Studio     works        evidence       reproduce\n</code></pre> <p>MEIDNet learns one latent space shared by crystal structures and their properties, then searches it for new materials that hit property targets while obeying the chemical and structural rules of a material family. The published cubic-perovskite model (band gap + formation enthalpy) is one application of it; the framework accepts any table of structures and scalar properties and any prototype family you describe.</p>      Your browser cannot play this video: download it.    MEIDNet in one minute: data, model, family, rules, targets, search and candidates. The same tour opens   on a first visit to the live Studio (the \u201c\u25b6 How it works\u201d button replays it). <ul> <li> <p> Learn multimodality</p> <p>What a modality is, the five challenges of multimodal learning, contrastive learning with a playground, and an interactive map of materials modalities.</p> <p> Learn</p> </li> <li> <p> Choose an architecture</p> <p>Early fusion, late fusion, shared latent spaces, cross-attention and contrastive learning, and an advisor that recommends one for your data and goal.</p> <p> Architecture Atlas</p> </li> <li> <p> Try MEIDNet</p> <p>See what it does in 30 seconds: the live Studio needs nothing installed; <code>meidnet demo</code> runs on any laptop.</p> <p> 5-minute quickstart</p> </li> <li> <p> Use MEIDNet on my data</p> <p>A table with an id, properties and CIFs is all you need. Four commands take you from \"is my data usable?\" to candidates with explanations.</p> <p> Bring your own dataset</p> </li> <li> <p> Change what it does</p> <p>Targets, elements, rules, family \u2014 all in one YAML file, or by moving sliders in MEIDNet Studio and exporting the file.</p> <p> Recipes</p> </li> <li> <p> Understand it</p> <p>How the two modalities are aligned, how the search works, what the reports mean, and where the limits are.</p> <p> How MEIDNet works</p> </li> </ul>"},{"location":"index.html#the-workflow","title":"The workflow","text":"<pre><code>   Data \u2500\u2500\u25ba Model \u2500\u2500\u25ba Family \u2500\u2500\u25ba Rules \u2500\u2500\u25ba Targets \u2500\u2500\u25ba Search \u2500\u2500\u25ba Candidates\n   table    shared    prototype  hard      property   latent     CIFs + report\n   + CIFs   latent    + sites    checks    goals      optimisation\n</code></pre> <p>Each block is a setting in <code>meidnet.yaml</code>, a node in MEIDNet Studio \u2014 where changing one block immediately shows its effect on the ones after it \u2014 and a section in the HTML report each step writes.</p>"},{"location":"index.html#honest-scope","title":"Honest scope","text":"<ul> <li>Generation works for prototype families: a fixed arrangement of sites whose occupants and cell size are   chosen (ABX\u2083 perovskites, A\u2082BB\u2032X\u2086 double perovskites, and anything you describe the same way). MEIDNet does   not invent new atomic arrangements. Details and roadmap.</li> <li>Predicted properties are the model's estimates. The reports say when a prediction is an extrapolation.   Confirm candidates with DFT or experiment; <code>meidnet screen</code> is a first filter.</li> </ul>"},{"location":"index.html#paper","title":"Paper","text":"<p>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, MEIDNet: Multimodal generative AI framework for inverse materials design, npj Computational Materials (2026). Citation and benchmarks \u00b7 code as published (v1.0).</p>"},{"location":"benchmarks/index.html","title":"MEIDNet Benchmarks","text":"<p>Reproducible results for multimodal materials representation and inverse design, one leaderboard per dataset. Different datasets are different tasks, so there is no ranking across them: compare rows only within a page. The idea follows Matbench Discovery: fixed data, inspectable submissions, plots and an evidence level per row.</p> 3datasets5results2reproduced here1DFT or experiment"},{"location":"benchmarks/index.html#datasets","title":"Datasets","text":"dataset role results reproduced highest evidence Carbon-24 published: structure-representation generalization (not a multimodal benchmark) 0 0 \u2013 MP-20 published: structure-representation generalization (not a multimodal benchmark) 0 0 \u2013 Perov-5 the published multimodal benchmark: alignment, property reconstruction and inverse design 5 2 DFT validated"},{"location":"benchmarks/index.html#the-evidence-ladder","title":"The evidence ladder","text":"level means \u25cb\u25cb\u25cb\u25cb\u25cb generated the model produced it \u25cf\u25cb\u25cb\u25cb\u25cb ML filtered it passed the family's rules and the model's own checks \u25cf\u25cf\u25cb\u25cb\u25cb MLIP validated relaxed and found stable with a machine-learned potential (<code>meidnet screen</code>, MACE) \u25cf\u25cf\u25cf\u25cb\u25cb DFT validated confirmed by density-functional calculations \u25cf\u25cf\u25cf\u25cf\u25cb experimentally validated synthesised and measured results per evidence level (all datasets)generated3MLIP validated1DFT validated1"},{"location":"benchmarks/index.html#status-of-a-row","title":"Status of a row","text":"<p>Published = quoted from the paper \u00b7 Community submitted = a pull request, validated by CI \u00b7 MEIDNet verified = reproduced here from the submitted configuration (<code>python scripts/benchmarks.py reproduce &lt;id&gt;</code>). How to contribute \u00b7 Databases by application</p>"},{"location":"benchmarks/index.html#latest-results","title":"Latest results","text":"date result dataset category evidence status 2026-10-03 MLIP screening of the shipped paper candidates (MACE-MP-0) Perov-5 Validation level \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated Community submitted 2026-10-03 Shipped checkpoint re-evaluated on the training split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-10-03 Shipped checkpoint re-evaluated on the held-out validation split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-05-29 MEIDNet (paper): inverse design of perovskites from property targets Perov-5 Conditional design \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated Published (from the paper) 2026-05-29 MEIDNet (paper): alignment of structure and property latents on Perov-5 Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated Published (from the paper)"},{"location":"benchmarks/carbon24.html","title":"Carbon-24","text":"<p>About 10,000 carbon structures with up to 24 atoms per cell (CDVAE split). In the MEIDNet paper it is the second structure-representation generalization test; no property modality and no checkpoint here.</p> <p>10,153 rows \u00b7 split: CDVAE: 6,091 train / 2,031 val / 2,031 test \u00b7 role in the paper: published: structure-representation generalization (not a multimodal benchmark) \u00b7 dataset source</p> 0results0reproduced here0categories\u2013highest evidence <p>Read a row</p> <p>Rows are compared only within this dataset and category; click a column header to sort. The ladder \u25cf\u25cb\u25cb\u25cb\u25cb \u2026 \u25cf\u25cf\u25cf\u25cf\u25cf is the evidence level (generated \u2192 ML filtered \u2192 MLIP validated \u2192 DFT validated \u2192 experimentally validated). Published = taken from the paper; MEIDNet verified = reproduced here from the submitted configuration.</p> <p>No results yet. Contribute one.</p>"},{"location":"benchmarks/mp20.html","title":"MP-20","text":"<p>About 45,000 Materials Project structures with at most 20 atoms per cell (CDVAE split). In the MEIDNet paper it tests how the crystal encoder and decoder generalise beyond perovskites; no property modality was aligned on it and no MP-20 checkpoint is shipped in this repository.</p> <p>45,231 rows \u00b7 split: CDVAE: 27,136 train / 9,047 val / 9,046 test \u00b7 role in the paper: published: structure-representation generalization (not a multimodal benchmark) \u00b7 dataset source</p> 0results0reproduced here0categories\u2013highest evidence <p>Read a row</p> <p>Rows are compared only within this dataset and category; click a column header to sort. The ladder \u25cf\u25cb\u25cb\u25cb\u25cb \u2026 \u25cf\u25cf\u25cf\u25cf\u25cf is the evidence level (generated \u2192 ML filtered \u2192 MLIP validated \u2192 DFT validated \u2192 experimentally validated). Published = taken from the paper; MEIDNet verified = reproduced here from the submitted configuration.</p> <p>No results yet. Contribute one.</p>"},{"location":"benchmarks/perov5.html","title":"Perov-5","text":"<p>18,928 cubic ABX3 perovskites with relaxed structures, formation enthalpy (heat_all, eV/atom) and direct band gap (dir_gap, eV). The CDVAE split of the Castelli et al. (2012) perovskite set; the dataset behind the published MEIDNet model and the live Studio.</p> <p>18,928 rows \u00b7 split: CDVAE: 11,356 train / 3,787 val / 3,785 test \u00b7 properties: <code>heat_all</code>, <code>dir_gap</code> \u00b7 role in the paper: the published multimodal benchmark: alignment, property reconstruction and inverse design \u00b7 dataset source</p> 5results2reproduced here3categoriesDFT validatedhighest evidence <p>Read a row</p> <p>Rows are compared only within this dataset and category; click a column header to sort. The ladder \u25cf\u25cb\u25cb\u25cb\u25cb \u2026 \u25cf\u25cf\u25cf\u25cf\u25cf is the evidence level (generated \u2192 ML filtered \u2192 MLIP validated \u2192 DFT validated \u2192 experimentally validated). Published = taken from the paper; MEIDNet verified = reproduced here from the submitted configuration.</p>"},{"location":"benchmarks/perov5.html#representation-quality","title":"Representation quality","text":"result split inputs cosine_matched l2_matched retrieval_top1 retrieval_top5 mae_heat_all r2_heat_all mae_dir_gap r2_dir_gap n_evaluated evidence status MEIDNet (paper): alignment of structure and property la\u2026 validation structure + 2 properties 0.97 0.24 \u2013 \u2013 \u2013 \u2013 \u2013 \u2013 \u2013 \u25cf\u25cb\u25cb\u25cb\u25cb generated Published (from the paper) Shipped checkpoint re-evaluated on the training split (\u2026 train structure + 2 properties 0.771 0.676 0.0101 0.0405 0.397 eV/atom 0.486 2.78 eV -29.8 11,356 materials \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified Shipped checkpoint re-evaluated on the held-out validat\u2026 val structure + 2 properties 0.771 0.676 0.0253 0.0908 0.392 eV/atom 0.497 2.79 eV -35.8 3,787 materials \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified How to read the representation quality table <ul> <li><code>cosine_matched</code> \u2014 mean cosine similarity between the structure and property latents of the same material (1 = aligned)</li> <li><code>l2_matched</code> \u2014 mean L2 distance between those two latents (0 = identical) (lower is better)</li> <li><code>retrieval_top1</code> \u2014 fraction of validation materials whose property latent is nearest to its own structure latent</li> <li><code>retrieval_top5</code> \u2014 the same within the five nearest</li> <li><code>mae_heat_all</code> \u2014 mean absolute error of <code>heat_all</code> predicted from the structure alone (physical units) (lower is better)</li> <li><code>r2_heat_all</code> \u2014 coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect</li> <li><code>mae_dir_gap</code> \u2014 mean absolute error of <code>dir_gap</code> predicted from the structure alone (physical units) (lower is better)</li> <li><code>r2_dir_gap</code> \u2014 coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect</li> <li><code>n_evaluated</code> \u2014 materials in the evaluation split</li> </ul> <p>Representation rows describe how well the shared latent space holds structures and properties; they say nothing about whether a generated material is real.</p> cosine_matched (higher is better)MEIDNet (paper): alignment of structur0.97Shipped checkpoint re-evaluated on the0.771Shipped checkpoint re-evaluated on the0.771retrieval_top1 (higher is better)Shipped checkpoint re-evaluated on the0.0253Shipped checkpoint re-evaluated on the0.0101mae_heat_all (lower is better)Shipped checkpoint re-evaluated on the0.397Shipped checkpoint re-evaluated on the0.392mae_dir_gap (lower is better)Shipped checkpoint re-evaluated on the2.79Shipped checkpoint re-evaluated on the2.78dir_gap: predicted from the structure vs. true (3,787 validation materials)02460246true dir_gap (eV)predicted dir_gap heat_all: predicted from the structure vs. true (3,787 validation materials)024024true heat_all (eV/atom)predicted heat_all cosine similarity of the two latents, per material0.650.70.750.80.850200400600800cosine(structure latent, property latent)count0.636\u20130.642: 10.672\u20130.678: 10.689\u20130.695: 10.695\u20130.701: 10.701\u20130.707: 20.707\u20130.713: 30.713\u20130.719: 30.719\u20130.725: 60.725\u20130.731: 50.731\u20130.737: 160.737\u20130.743: 280.743\u20130.748: 520.748\u20130.754: 1080.754\u20130.76: 2970.76\u20130.766: 7360.766\u20130.772: 7540.772\u20130.778: 5820.778\u20130.784: 7710.784\u20130.79: 2510.79\u20130.796: 640.796\u20130.802: 290.802\u20130.808: 290.808\u20130.813: 180.813\u20130.819: 170.819\u20130.825: 50.825\u20130.831: 30.831\u20130.837: 30.855\u20130.861: 1mean"},{"location":"benchmarks/perov5.html#conditional-design","title":"Conditional design","text":"result split inputs n_generated n_sun sun_rate evidence status MEIDNet (paper): inverse design of perovskites from pro\u2026 \u2013 structure + 2 properties 140 structures 19 structures 0.136 \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated Published (from the paper) How to read the conditional design table <ul> <li><code>n_generated</code> \u2014 candidates the search produced</li> <li><code>n_sun</code> \u2014 stable, unique and novel candidates</li> <li><code>sun_rate</code> \u2014 n_sun / n_generated</li> </ul> <p>Design rows are funnels: every number after <code>n_generated</code> is a subset of the one before. A high SUN rate on few candidates is weaker evidence than a lower rate on many.</p> MEIDNet (paper): inverse design of perovskites from property targets: from generated to validatedgenerated140stable, unique, novel19"},{"location":"benchmarks/perov5.html#validation-level","title":"Validation level","text":"result split inputs n_screened evidence status MLIP screening of the shipped paper candidates (MACE-MP\u2026 \u2013 structure + 2 properties 26 structures \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated Community submitted How to read the validation level table <ul> <li><code>n_screened</code> \u2014 structures relaxed with the MLIP</li> </ul> <p>Validation rows count structures that survived each check. <code>n_stable</code> depends on the stability threshold used (0.10 eV/atom above the hull by default).</p>"},{"location":"benchmarks/perov5.html#insights","title":"Insights","text":"<ul> <li>Best <code>cosine_matched</code>: MEIDNet (paper): alignment of structure and property latents on Perov-5 (0.97), published (from the paper).</li> <li>Best <code>l2_matched</code>: MEIDNet (paper): alignment of structure and property latents on Perov-5 (0.24), published (from the paper).</li> <li>Best <code>retrieval_top1</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.0253), meidnet verified.</li> <li>Best <code>retrieval_top5</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.0908), meidnet verified.</li> <li>Best <code>mae_heat_all</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.392 eV/atom), meidnet verified.</li> <li>Best <code>r2_heat_all</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.497), meidnet verified.</li> <li>Best <code>mae_dir_gap</code>: Shipped checkpoint re-evaluated on the training split (this code) (2.78 eV), meidnet verified.</li> <li>Best <code>r2_dir_gap</code>: Shipped checkpoint re-evaluated on the training split (this code) (-29.8), meidnet verified.</li> <li>Evidence levels present: generated, MLIP validated, DFT validated. 2 of 5 rows have been reproduced here.</li> <li>2 row(s) are quoted from the paper; a verified re-run of the same checkpoint, where it exists, is the row to trust for the exact numbers of this code version.</li> <li>A re-run of MEIDNet (paper): alignment of structure and property latents on Perov-5 here did not agree (<code>cosine_matched</code> 0.97 \u2192 0.771; <code>l2_matched</code> 0.24 \u2192 0.676). The row keeps its status; see its page for the full comparison.</li> </ul>"},{"location":"benchmarks/results/paper-inverse-design.html","title":"MEIDNet (paper): inverse design of perovskites from property targets","text":"<p>Published (from the paper) \u00b7 \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 1.0 (paper), perovskite_abx3) Category Conditional design Submitted by Anand Babu (UCLouvain) on 2026-05-29 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth"},{"location":"benchmarks/results/paper-inverse-design.html#results","title":"Results","text":"metric value split meaning note <code>n_generated</code> 140 structures candidates the search produced candidates generated from property targets in the campaign reported in the paper <code>n_sun</code> 19 structures stable, unique and novel candidates stable, unique and novel after screening <code>sun_rate</code> 0.136 n_sun / n_generated 19 of 140 MEIDNet (paper): inverse design of perovskites from property targets: from generated to validatedgenerated140stable, unique, novel19"},{"location":"benchmarks/results/paper-inverse-design.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated.</p> <p>Evidence:</p> <ul> <li>https://doi.org/10.1038/s41524-026-02153-3</li> <li><code>examples/perov5/paper_results/</code></li> </ul> <p>These numbers are quoted from the paper.</p>"},{"location":"benchmarks/results/paper-inverse-design.html#notes","title":"Notes","text":"<p>Numbers quoted from the paper. The 26 CIFs in examples/perov5/paper_results/ are the candidates shipped with the paper; the MLIP screening row re-runs <code>meidnet screen</code> on them.</p>"},{"location":"benchmarks/results/paper-representation.html","title":"MEIDNet (paper): alignment of structure and property latents on Perov-5","text":"<p>Published (from the paper) \u00b7 \u25cf\u25cb\u25cb\u25cb\u25cb generated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 1.0 (paper), perovskite_abx3) Category Representation quality Submitted by Anand Babu (UCLouvain) on 2026-05-29 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Notes on the model 128-dimensional shared latent space; 2000 epochs, contrastive warm-up over the first 1200"},{"location":"benchmarks/results/paper-representation.html#results","title":"Results","text":"metric value split meaning note <code>cosine_matched</code> 0.97 validation mean cosine similarity between the structure and property latents of the same material (1 = aligned) mean cosine similarity between the structure and property latents of the same material, as reported in the paper <code>l2_matched</code> 0.24 validation mean L2 distance between those two latents (0 = identical) mean L2 distance between the two latents of the same material, as reported in the paper"},{"location":"benchmarks/results/paper-representation.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cb\u25cb\u25cb\u25cb generated.</p> <p>Evidence:</p> <ul> <li>https://doi.org/10.1038/s41524-026-02153-3</li> </ul>"},{"location":"benchmarks/results/paper-representation.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>, split val:</p> <pre><code>python scripts/benchmarks.py reproduce paper-representation\n</code></pre> <p>These numbers are quoted from the paper.</p>"},{"location":"benchmarks/results/paper-representation.html#re-run-here-did-not-agree","title":"Re-run here did not agree","text":"<p>On 2026-10-03 this repository's code re-ran the recipe above (AMD64 CPU, MEIDNet 2.0.0) and obtained different numbers. The row keeps its status; the comparison is shown so that readers can judge it. Possible reasons: a different split or protocol in the original measurement, a different definition of the metric, or a checkpoint that is not the one the numbers were measured on.</p> metric this row obtained here agree <code>cosine_matched</code> 0.97 0.771 no <code>l2_matched</code> 0.24 0.676 no <p>Also obtained in that run: <code>mae_heat_all</code> = 0.392, <code>r2_heat_all</code> = 0.497, <code>mae_dir_gap</code> = 2.79, <code>r2_dir_gap</code> = -35.8, <code>retrieval_top1</code> = 0.0253, <code>retrieval_top5</code> = 0.0908, <code>n_evaluated</code> = 3,787</p>"},{"location":"benchmarks/results/paper-representation.html#notes","title":"Notes","text":"<p>Numbers quoted from the paper (npj Computational Materials, 2026). <code>reproduce</code> re-evaluates the shipped checkpoint on the CDVAE validation split and adds MAE, R2 and retrieval metrics; cosine_matched is compared with the quoted value.</p>"},{"location":"benchmarks/results/paper-results-mace-screen.html","title":"MLIP screening of the shipped paper candidates (MACE-MP-0)","text":"<p>Community submitted \u00b7 \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 2.0.0, perovskite_abx3) Category Validation level Submitted by Anand Babu (UCLouvain) on 2026-10-03 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Resources laptop CPU"},{"location":"benchmarks/results/paper-results-mace-screen.html#results","title":"Results","text":"metric value split meaning note <code>n_screened</code> 26 structures structures relaxed with the MLIP the CIFs in examples/perov5/paper_results/"},{"location":"benchmarks/results/paper-results-mace-screen.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated.</p> <p>Evidence:</p> <ul> <li><code>examples/perov5/paper_results/</code></li> </ul>"},{"location":"benchmarks/results/paper-results-mace-screen.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>:</p> <pre><code>meidnet screen examples/perov5/paper_results --train-csv data/perov5/train.csv\n</code></pre>"},{"location":"benchmarks/results/paper-results-mace-screen.html#notes","title":"Notes","text":"<p>n_stable / n_unique / n_novel / n_sun are filled in by <code>python scripts/benchmarks.py reproduce paper-results-mace-screen</code> (MACE-MP-0 relaxation, about ten minutes on a CPU).</p>"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html","title":"Shipped checkpoint re-evaluated on the training split (this code)","text":"<p>MEIDNet verified \u00b7 \u25cf\u25cb\u25cb\u25cb\u25cb generated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 2.0.0, perovskite_abx3) Category Representation quality Submitted by Anand Babu (UCLouvain) on 2026-10-03 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Resources laptop CPU (AMD64)"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#results","title":"Results","text":"metric value split meaning note <code>cosine_matched</code> 0.771 train mean cosine similarity between the structure and property latents of the same material (1 = aligned) mean over the 11,356 training materials <code>l2_matched</code> 0.676 train mean L2 distance between those two latents (0 = identical) <code>retrieval_top1</code> 0.0101 train fraction of validation materials whose property latent is nearest to its own structure latent <code>retrieval_top5</code> 0.0405 train the same within the five nearest <code>mae_heat_all</code> 0.397 eV/atom train mean absolute error of <code>heat_all</code> predicted from the structure alone (physical units) <code>r2_heat_all</code> 0.486 train coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>mae_dir_gap</code> 2.78 eV train mean absolute error of <code>dir_gap</code> predicted from the structure alone (physical units) <code>r2_dir_gap</code> -29.8 train coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>n_evaluated</code> 11,356 materials train materials in the evaluation split"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cb\u25cb\u25cb\u25cb generated.</p> <p>Evidence:</p> <ul> <li><code>checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth</code></li> </ul>"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>, split train:</p> <pre><code>python scripts/benchmarks.py reproduce shipped-checkpoint-train-split\n</code></pre>"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#verified-here","title":"Verified here","text":"<p>Reproduced on 2026-10-03 with MEIDNet 2.0.0 by Anand Babu on AMD64 CPU:</p> metric submitted obtained here <code>mae_heat_all</code> 0.397 0.397 <code>r2_heat_all</code> 0.486 0.486 <code>mae_dir_gap</code> 2.78 2.78 <code>r2_dir_gap</code> -29.8 -29.8 <code>retrieval_top1</code> 0.0101 0.0101 <code>retrieval_top5</code> 0.0405 0.0405 <code>cosine_matched</code> 0.771 0.771 <code>n_evaluated</code> 11,356 11,356 <code>l2_matched</code> 0.676 0.676"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#per-material-analysis","title":"Per-material analysis","text":"<p>Every validation material, predicted from its structure alone; the dashed line is perfect agreement.</p> dir_gap: predicted from the structure vs. true (11,356 validation materials)0246802468true dir_gap (eV)predicted dir_gap heat_all: predicted from the structure vs. true (11,356 validation materials)024024true heat_all (eV/atom)predicted heat_all cosine similarity of the two latents, per material00.502,5005,0007,50010,000cosine(structure latent, property latent)count-0.436\u2013-0.4: 10.211\u20130.247: 20.391\u20130.427: 10.463\u20130.499: 10.499\u20130.535: 10.535\u20130.571: 10.607\u20130.643: 30.643\u20130.679: 20.679\u20130.715: 160.715\u20130.751: 4940.751\u20130.787: 101520.787\u20130.823: 6270.823\u20130.859: 480.859\u20130.895: 50.895\u20130.931: 2mean"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#notes","title":"Notes","text":"<p>The same evaluation on the data the model was trained on. The numbers equal the held-out ones within noise: the gap between the paper's figures and these is not a train / test effect.</p>"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html","title":"Shipped checkpoint re-evaluated on the held-out validation split (this code)","text":"<p>MEIDNet verified \u00b7 \u25cf\u25cb\u25cb\u25cb\u25cb generated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 2.0.0, perovskite_abx3) Category Representation quality Submitted by Anand Babu (UCLouvain) on 2026-10-03 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Resources laptop CPU (AMD64) Notes on the model the checkpoint shipped with the repository and used by the live Studio; 128-dimensional shared latent space"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#results","title":"Results","text":"metric value split meaning note <code>cosine_matched</code> 0.771 val mean cosine similarity between the structure and property latents of the same material (1 = aligned) mean over the 3,787 validation materials <code>l2_matched</code> 0.676 val mean L2 distance between those two latents (0 = identical) unit latents: sqrt(2 - 2 cos) <code>retrieval_top1</code> 0.0253 val fraction of validation materials whose property latent is nearest to its own structure latent <code>retrieval_top5</code> 0.0908 val the same within the five nearest <code>mae_heat_all</code> 0.392 eV/atom val mean absolute error of <code>heat_all</code> predicted from the structure alone (physical units) predicted from the structure latent alone <code>r2_heat_all</code> 0.497 val coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>mae_dir_gap</code> 2.79 eV val mean absolute error of <code>dir_gap</code> predicted from the structure alone (physical units) 96% of the rows have a direct gap of 0 eV; the structure-only predictor does not capture that <code>r2_dir_gap</code> -35.8 val coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>n_evaluated</code> 3,787 materials val materials in the evaluation split"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cb\u25cb\u25cb\u25cb generated.</p> <p>Evidence:</p> <ul> <li><code>checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth</code></li> <li><code>benchmarks/verified/shipped-checkpoint-val-split.predictions.csv</code></li> </ul>"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>, split val:</p> <pre><code>python scripts/benchmarks.py reproduce shipped-checkpoint-val-split\n</code></pre>"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#verified-here","title":"Verified here","text":"<p>Reproduced on 2026-10-03 with MEIDNet 2.0.0 by Anand Babu on AMD64 CPU:</p> metric submitted obtained here <code>mae_heat_all</code> 0.392 0.392 <code>r2_heat_all</code> 0.497 0.497 <code>mae_dir_gap</code> 2.79 2.79 <code>r2_dir_gap</code> -35.8 -35.8 <code>retrieval_top1</code> 0.0253 0.0253 <code>retrieval_top5</code> 0.0908 0.0908 <code>cosine_matched</code> 0.771 0.771 <code>n_evaluated</code> 3,787 3,787 <code>l2_matched</code> 0.676 0.676"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#per-material-analysis","title":"Per-material analysis","text":"<p>Every validation material, predicted from its structure alone; the dashed line is perfect agreement.</p> dir_gap: predicted from the structure vs. true (3,787 validation materials)02460246true dir_gap (eV)predicted dir_gap heat_all: predicted from the structure vs. true (3,787 validation materials)024024true heat_all (eV/atom)predicted heat_all cosine similarity of the two latents, per material0.650.70.750.80.850200400600800cosine(structure latent, property latent)count0.636\u20130.642: 10.672\u20130.678: 10.689\u20130.695: 10.695\u20130.701: 10.701\u20130.707: 20.707\u20130.713: 30.713\u20130.719: 30.719\u20130.725: 60.725\u20130.731: 50.731\u20130.737: 160.737\u20130.743: 280.743\u20130.748: 520.748\u20130.754: 1080.754\u20130.76: 2970.76\u20130.766: 7360.766\u20130.772: 7540.772\u20130.778: 5820.778\u20130.784: 7710.784\u20130.79: 2510.79\u20130.796: 640.796\u20130.802: 290.802\u20130.808: 290.808\u20130.813: 180.813\u20130.819: 170.819\u20130.825: 50.825\u20130.831: 30.831\u20130.837: 30.855\u20130.861: 1mean"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#notes","title":"Notes","text":"<p>What this code measures on the shipped checkpoint, with every prediction saved (parity plots below). The paper's representation row quotes a cosine of about 0.97; this re-run does not reach it, see the comparison on that row's page.</p>"},{"location":"community/index.html","title":"Community results","text":"<p>Every submitted result, newest first. Contribute yours.</p> date result dataset category evidence status 2026-10-03 MLIP screening of the shipped paper candidates (MACE-MP-0) Perov-5 Validation level \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated Community submitted 2026-10-03 Shipped checkpoint re-evaluated on the training split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-10-03 Shipped checkpoint re-evaluated on the held-out validation split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-05-29 MEIDNet (paper): inverse design of perovskites from property targets Perov-5 Conditional design \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated Published (from the paper) 2026-05-29 MEIDNet (paper): alignment of structure and property latents on Perov-5 Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated Published (from the paper)"},{"location":"community/contribute.html","title":"Contribute a benchmark result","text":"<p>MEIDNet Benchmarks collect results per dataset, each with the evidence behind it. Anyone who runs MEIDNet on a dataset \u2014 the published Perov-5, or their own \u2014 can submit a result. Submissions go through GitHub, so every number has a public record and a reviewer.</p>"},{"location":"community/contribute.html#what-a-submission-is","title":"What a submission is","text":"<p>One JSON file, <code>benchmarks/submissions/&lt;dataset&gt;/&lt;id&gt;.json</code>, that follows <code>benchmarks/schema.json</code>. The short version:</p> field what to put <code>dataset</code> an id from <code>benchmarks/datasets/</code> (<code>perov5</code>, \u2026); a new dataset needs its own <code>datasets/&lt;id&gt;.json</code> with source, size and split <code>modalities</code> what the model aligned: <code>structure</code>, <code>property:&lt;column&gt;</code> (one per property), later <code>xrd</code>, \u2026 <code>model</code> MEIDNet version, family and variant, the configuration (<code>meidnet.yaml</code>), checkpoint link and its SHA-256 <code>category</code> <code>representation_quality</code>, <code>conditional_design</code>, <code>efficiency</code> or <code>validation</code> \u2014 one category per file <code>results</code> metric \u2192 <code>{value, unit}</code>; only the metric names allowed for that category (listed in the schema) <code>validation_level</code> the evidence ladder: <code>generated</code> \u2192 <code>ml_filtered</code> \u2192 <code>mlip_validated</code> \u2192 <code>dft_validated</code> \u2192 <code>experimentally_validated</code> <code>evidence</code> public links: CIF archive, <code>stability.csv</code>, DFT inputs/outputs, paper DOI <code>reproduce</code> configuration, seed and the exact command that produces the numbers <code>status</code> always <code>community_submitted</code> for a new file"},{"location":"community/contribute.html#the-five-things-people-get-wrong","title":"The five things people get wrong","text":"<ol> <li>Mixing categories in one file. A representation result (MAE, R\u00b2, retrieval) and a design result (SUN rate)    are two files.</li> <li>A metric name that is not in the list. <code>validate</code> refuses it; the lists exist so that rows are comparable.</li> <li>No split. State the split you evaluated on (or link a file of ids). A number on an unknown split is not comparable.</li> <li>Evidence that is not public. A result without an inspectable artefact stays at <code>generated</code>.</li> <li>Claiming a validation level the evidence does not support. <code>dft_validated</code> needs the DFT outputs, not the intention.</li> </ol>"},{"location":"community/contribute.html#how-it-works","title":"How it works","text":"<ol> <li>Copy <code>benchmarks/submissions/_template.json</code>, fill it in, run <code>python scripts/benchmarks.py validate</code>.</li> <li>Open a pull request (the template has a checklist). CI runs <code>validate</code>.</li> <li>Once merged, the result is rendered on the dataset's benchmark page and gets a permanent page of its own,    marked Community submitted.</li> <li>The maintainer may run <code>python scripts/benchmarks.py reproduce &lt;id&gt;</code>: it re-runs the stated command from the    stated configuration and compares the numbers. If they agree within the tolerances, the row becomes    MEIDNet verified (a record in <code>benchmarks/verified/</code>, which only the maintainer can add).</li> </ol> <p>Verified means reproduced from the submitted configuration with this code. It does not mean the material is real: that is what the validation level says. A re-run that does not agree is kept too (<code>benchmarks/verified/&lt;id&gt;.attempt.json</code>) and shown on the row's page next to the submitted numbers, so that a disagreement is visible rather than silent \u2014 this applies to the paper's own rows as much as to anyone else's.</p> <p>Representation rows that are reproduced here also get their per-material predictions saved (<code>benchmarks/verified/&lt;id&gt;.predictions.csv</code>), which the pages turn into parity plots and a latent-agreement histogram.</p> <p>If you prefer not to open a pull request, the issue template Benchmark submission asks for the same fields and the maintainer converts it. Submissions are reviewed by the author of MEIDNet; there is no promised response time.</p>"},{"location":"examples/double-perovskite.html","title":"Double perovskites A\u2082BB\u2032X\u2086","text":"<p><code>meidnet/families/double_perovskite_a2bbx6.yaml</code> is a second family written with the same ingredients as the perovskite one. It shows what a family file can express:</p> <ul> <li>a 10-site prototype in the rhombohedral primitive cell of an fcc lattice (<code>lattice: fcc_primitive</code>:   edge a/\u221a2, angles 60\u00b0), A at (\u00bc,\u00bc,\u00bc) and (\u00be,\u00be,\u00be), B at the origin, B\u2032 at (\u00bd,\u00bd,\u00bd), six X;</li> <li>four site groups with their own element lists and oxidation states (B: +1/+2, B\u2032: +3/+4);</li> <li>a cell-size rule <code>a = 2 (r_B + r_B\u2032 + 2 r_X)</code> (one B\u2013X\u2013B\u2032 chain per half cell edge);</li> <li>rules that average radii over several groups: <code>tolerance_factor: {A: A, B: [B1, B2], X: X}</code>.</li> </ul> <pre><code>meidnet families double_perovskite_a2bbx6 --variant halide\n</code></pre> <pre><code>Rock-salt ordered double perovskite A2BB'X6 (Fm-3m)  [variant: halide]\n  formula pattern: A2B1B2X6   (10 atoms per cell)\n  A: 9 candidate elements - Large cation in the cuboctahedral cavity\n  B1: 15 candidate elements - First octahedral cation (usually +1 or +2)\n  B2: 20 candidate elements - Second octahedral cation (usually +3 or +4)\n  X: 4 candidate elements - Anion shared between neighbouring B and B' octahedra\n</code></pre> <p>The halide variant has 10,800 compositions; 305 pass every rule (Cs\u2082AgBiI\u2086 among them). Candidates are built in the 10-atom primitive cell; the symmetry refinement step writes them out as the conventional 40-atom Fm-3m cell, which is what the CIF files contain.</p>"},{"location":"examples/double-perovskite.html#using-it","title":"Using it","text":"<p>A model for this family must be trained on double-perovskite structures aligned to the prototype:</p> <pre><code>family: double_perovskite_a2bbx6\ndata:\n  table: double_perovskites.csv\n  properties: [{column: band_gap, unit: eV}]\n  align_to_prototype: true\ngeneration:\n  family: double_perovskite_a2bbx6\n  variant: halide\n  objectives: [{property: band_gap}]\n  targets: [{band_gap: 2.0}]\n</code></pre> <p>The published checkpoint</p> <p>The Perov-5 model was trained on 5-atom ABX\u2083 cells. The Studio will score double-perovskite compositions with it (the encoder accepts any cell up to 20 atoms), but those predictions are far outside its training distribution \u2014 a demonstration of the mechanics, not a result.</p>"},{"location":"examples/perov5.html","title":"Perov-5 \u2014 the paper's experiment","text":"<p>The published MEIDNet model: cubic ABX\u2083 perovskites from the Perov-5 dataset (CDVAE split of Castelli et al.), properties formation enthalpy (<code>heat_all</code>, eV/atom) and direct band gap (<code>dir_gap</code>, eV).</p>"},{"location":"examples/perov5.html#use-the-published-checkpoint","title":"Use the published checkpoint","text":"<pre><code>meidnet demo --family halide --band-gap 2.0            # or: oxide, chalcogenide, nitride\nmeidnet studio                                          # explore it live\n</code></pre> <p>The three checkpoints from the paper are in <code>checkpoints/</code>; <code>..._propertyaware_2k.pth</code> is the recommended one. <code>meidnet info checkpoints/&lt;file&gt;</code> describes any of them. They load into MEIDNet 2 unchanged.</p>"},{"location":"examples/perov5.html#retrain-from-scratch","title":"Retrain from scratch","text":"<pre><code>meidnet download-data                       # train/val/test CSVs \u2192 data/perov5/ (18,928 rows, 18 MB)\nmeidnet train examples/perov5/meidnet.yaml  # 200 epochs \u2248 1 h on a laptop GPU\nmeidnet generate examples/perov5/meidnet.yaml\n</code></pre> <p><code>examples/perov5/meidnet.yaml</code> reproduces the paper's settings, including two that differ from MEIDNet 2's defaults: <code>normalize: false</code> for both properties and <code>align_to_prototype: false</code> (atoms in file order), and the 1200-epoch alignment ramp.</p>"},{"location":"examples/perov5.html#the-papers-generated-structures","title":"The paper's generated structures","text":"<p><code>examples/perov5/paper_results/</code> holds the 26 CIFs (halide, oxide, chalcogenide) reported in the paper. <code>meidnet screen examples/perov5/paper_results --train-csv data/perov5/train.csv</code> recomputes their MACE stability.</p>"},{"location":"examples/perov5.html#regression-against-v1","title":"Regression against v1","text":"<p><code>tests/legacy_v1/</code> contains the v1.0 code byte for byte. <code>pytest -m slow</code> runs it and MEIDNet 2 on the same targets and asserts identical CIFs, predictions and file names.</p>"},{"location":"explore/capabilities.html","title":"Capabilities: what works today","text":"<p>Generated from the code, so it cannot promise more than the code does. Supported = in this release; Community-tested = a benchmark result reproduced here names it; Planned = designed for, not implemented (roadmap).</p>"},{"location":"explore/capabilities.html#modalities","title":"Modalities","text":"modality status how Crystal structure (CIF, up to <code>max_sites</code> atoms, default 20) Community-tested an equivariant graph encoder; structures are aligned to the family prototype Scalar properties (any number of numeric columns) Community-tested one property encoder; every column becomes a target you can set Published properties: direct band gap (<code>dir_gap</code>), formation enthalpy (<code>heat_all</code>) Community-tested the shipped Perov-5 checkpoint Vector modalities: binned XRD, DOS Planned an encoder per vector modality into the shared latent space Spectra (Raman, UV-Vis) Planned as vector modalities Text (descriptions, synthesis) Planned a text encoder into the shared latent space Images (microscopy) Planned an image encoder into the shared latent space"},{"location":"explore/capabilities.html#fusion","title":"Fusion","text":"<p>One scheme: early fusion - the structure latent and the property latent of a material are averaged into the joint latent (<code>meidnet/model.py</code>), after a contrastive alignment whose weight ramps up over <code>training.contrastive_warmup_epochs</code> (the curriculum of the paper). There is no late-fusion option; a selector would be a fiction.</p>"},{"location":"explore/capabilities.html#input-formats","title":"Input formats","text":"format note CSV any delimiter pandas reads Excel (.xlsx, .xls) first sheet JSON table records Parquet needs the optional pyarrow CIF text in a column one structure per row CIF files (.cif, or a .zip of them) named .cif"},{"location":"explore/capabilities.html#material-families","title":"Material families","text":"family variants site groups <code>double_perovskite_a2bbx6</code> - Rock-salt ordered double perovskite A2BB'X6 (Fm-3m) <code>halide</code>, <code>oxide</code> A, B1, B2, X <code>perovskite_abx3</code> - Cubic ABX3 perovskite (Pm-3m) <code>oxide</code>, <code>halide</code>, <code>chalcogenide</code>, <code>nitride</code> A, B, X <p>Any other prototype family can be described in a family file.</p>"},{"location":"explore/capabilities.html#rules-hard-constraints","title":"Rules (hard constraints)","text":"<ul> <li><code>bond_window</code> - Sensible {from}\u2013{to} bonds</li> <li><code>charge_neutrality</code> - Charge balance</li> <li><code>min_distance</code> - No overlapping atoms</li> <li><code>octahedral_factor</code> - Octahedral factor</li> <li><code>property_window</code> - Predicted {property} window</li> <li><code>tolerance_factor</code> - Goldschmidt tolerance factor</li> </ul>"},{"location":"explore/capabilities.html#search-terms-soft","title":"Search terms (soft)","text":"<ul> <li><code>cubic_cell</code> - Pulls the decoded cell towards a cube of the family's reference volume.</li> <li><code>entropy_bonus</code> - Rewards uncertainty early in a round so the search explores (fades to zero).</li> <li><code>group_consistency</code> - Sites of the same group should agree on their element (ramps up during a round).</li> <li><code>prototype_alignment</code> - Pulls decoded positions onto the prototype's site positions.</li> <li><code>short_distance_penalty</code> - Large penalty if any two decoded atoms come closer than <code>min</code>.</li> <li><code>site_repulsion</code> - Short-range exponential repulsion between decoded atoms.</li> <li><code>site_separation</code> - Keeps decoded atoms from collapsing onto each other (1/distance).</li> <li><code>soft_charge</code> - Expected total charge of the most likely composition should be zero (ramps up during a round).</li> <li><code>tolerance_penalty</code> - Soft version of the Goldschmidt tolerance-factor window, using the most likely elements.</li> </ul>"},{"location":"explore/capabilities.html#after-generation","title":"After generation","text":"<ul> <li><code>meidnet screen</code>: MACE-MP-0 relaxation and the stable / unique / novel (SUN) counts (optional <code>pip install meidnet[stability]</code>). DFT and experiments are yours: the benchmark evidence ladder records them.</li> </ul>"},{"location":"explore/databases.html","title":"Databases by application","text":"<p>Where to find training data and realistic targets: computed (DFT) databases give structures with properties, experimental databases give measured values and refined structures. MEIDNet needs, per material, one crystal structure plus one or more scalar properties (what data do I need?); the last column says how far each source is from that.</p> <p>Pick an application:</p> <ul> <li>Solar cells and photovoltaics</li> <li>Semiconductor physics</li> <li>Batteries and ionic conductors</li> <li>Thermoelectrics</li> <li>Catalysis and surfaces</li> <li>Magnetism and superconductivity</li> <li>Mechanical, dielectric and piezoelectric</li> <li>2D materials</li> <li>Porous materials and MOFs</li> <li>Generative-model benchmarks</li> <li>All databases</li> </ul>"},{"location":"explore/databases.html#solar-cells-and-photovoltaics","title":"Solar cells and photovoltaics","text":"<p>What to look for: band gap in the 1.1-1.8 eV window (single junction) or 1.6-2.0 eV (tandem top cell), direct gap, strong absorption, formation energy / hull distance, defect tolerance, device records for the experimental side</p> <p>As MEIDNet targets: dir_gap or any gap column + heat_all / e_above_hull; family = perovskite_abx3 (halide, oxide, chalcogenide), or your own prototype</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure NOMAD computed more than 10,000,000 calculations raw and normalised DFT outputs from many codes; band structures, DOS, energies; FAIR provenance needs curation: pick one code and functional before building a table C2DB computed about 4,000 monolayers stability (dynamic and thermodynamic), PBE / HSE / GW gaps, magnetic state, optical absorption, piezoelectric and Raman data for monolayers ready for a 2D prototype family Cubic perovskites (CMR, Castelli et al.) computed about 19,000 ABX3 compositions formation (heat of formation) energies and direct / indirect gaps (GLLB-SC) for cubic ABX3 with O, N, S, F anions and their mixtures - the source of Perov-5 the published MEIDNet benchmark (Perov-5 split): meidnet download-data matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Perovskite Database Project experimental more than 42,000 solar-cell devices device-level records: composition, architecture, efficiency, Voc, Jsc, FF, stability, processing conditions, from the literature no crystal structures per record: use it to choose realistic gap / composition targets and to check candidates against what has been made Hybrid organic-inorganic perovskites (Kim et al. 2017) computed 1,346 HOIPs DFT structures, band gaps and dielectric constants of ABX3 hybrid perovskites ready, but organic A-site cations need a family file with molecular site groups (not shipped) Experimental band gaps (Zhuo et al. 2018) experimental 6,354 compounds measured band gaps by composition, with the matbench_expt_gap subset as a standard task targets and validation of predicted gaps; no structures Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#semiconductor-physics","title":"Semiconductor physics","text":"<p>What to look for: band gaps at several levels of theory (PBE, HSE, mBJ) and experimental, effective masses, dielectric constants, carrier mobility proxies, band edges, phonons</p> <p>As MEIDNet targets: gap, effective mass or dielectric constant as scalar targets; keep the level of theory consistent inside one table</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) NOMAD computed more than 10,000,000 calculations raw and normalised DFT outputs from many codes; band structures, DOS, energies; FAIR provenance needs curation: pick one code and functional before building a table Materials Cloud computed curated archives (MC3D, MC2D, phonons, ...) MC3D relaxed structures, MC2D exfoliable monolayers, phonon database, Sssp pseudopotential sets, workflow provenance (AiiDA) ready for MC3D / MC2D tables C2DB computed about 4,000 monolayers stability (dynamic and thermodynamic), PBE / HSE / GW gaps, magnetic state, optical absorption, piezoelectric and Raman data for monolayers ready for a 2D prototype family Cubic perovskites (CMR, Castelli et al.) computed about 19,000 ABX3 compositions formation (heat of formation) energies and direct / indirect gaps (GLLB-SC) for cubic ABX3 with O, N, S, F anions and their mixtures - the source of Perov-5 the published MEIDNet benchmark (Perov-5 split): meidnet download-data Matbench and Matbench Discovery computed 13 tasks (Matbench); 257,000 WBM test structures (Discovery) standardised property-prediction tasks (gap, formation energy, moduli, dielectric, ...) with fixed folds; Discovery ranks universal potentials on stability prediction with F1, DAF, RMSD and a leaderboard with plots use the tasks as property tables; the Discovery leaderboard is the model for MEIDNet Benchmarks matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Hybrid organic-inorganic perovskites (Kim et al. 2017) computed 1,346 HOIPs DFT structures, band gaps and dielectric constants of ABX3 hybrid perovskites ready, but organic A-site cations need a family file with molecular site groups (not shipped) Experimental band gaps (Zhuo et al. 2018) experimental 6,354 compounds measured band gaps by composition, with the matbench_expt_gap subset as a standard task targets and validation of predicted gaps; no structures QMOF computed about 20,000 MOFs DFT-optimised MOF structures with band gaps and charges beyond today's cell-size limit; targets and descriptors Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#batteries-and-ionic-conductors","title":"Batteries and ionic conductors","text":"<p>What to look for: ionic conductivity, migration barriers, voltage, stability window, volume change, hull distance</p> <p>As MEIDNet targets: formation energy + conductivity (log scale) as targets; families with a mobile-ion site group</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Liverpool Ionic Conductivity Database (LiIonDB) experimental about 800 measurements experimental Li-ion conductivities with temperature and the reported phase, from the literature targets (log conductivity); join with COD / Materials Project structures Materials Project battery explorer computed about 4,000 intercalation electrodes average voltage, capacity, volume change and stability of intercalation electrodes computed from the MP structures ready: voltage and capacity as targets on the host-structure family Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#thermoelectrics","title":"Thermoelectrics","text":"<p>What to look for: Seebeck coefficient, electrical and thermal conductivity, zT, carrier concentration and temperature; experimental data is temperature-resolved</p> <p>As MEIDNet targets: zT or Seebeck at a fixed temperature as a scalar target</p> database kind size features use with MEIDNet AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets UCSB thermoelectrics (Gaultois et al.) experimental about 1,100 compounds experimental Seebeck, resistivity, thermal conductivity, zT at a given temperature targets; join with structures Starrydata experimental tens of thousands of digitised curves temperature-dependent thermoelectric and other property curves digitised from papers targets at a chosen temperature Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#catalysis-and-surfaces","title":"Catalysis and surfaces","text":"<p>What to look for: adsorption energies, reaction barriers, surface energies, work functions</p> <p>As MEIDNet targets: bulk descriptors (formation energy, d-band proxies) only - MEIDNet works on bulk prototypes, not slabs</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table NOMAD computed more than 10,000,000 calculations raw and normalised DFT outputs from many codes; band structures, DOS, energies; FAIR provenance needs curation: pick one code and functional before building a table Open Catalyst (OC20 / OC22) computed more than 1,300,000 relaxations adsorbate-surface relaxations with energies and forces; the standard catalysis ML benchmark surfaces are out of MEIDNet's scope; use the bulk subsets as structure sources Catalysis-Hub computed more than 100,000 reaction energies adsorption and reaction energies on surfaces with the DFT settings used descriptors only Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#magnetism-and-superconductivity","title":"Magnetism and superconductivity","text":"<p>What to look for: magnetic moment and ordering, Curie / Neel temperature, superconducting Tc with the structure it belongs to</p> <p>As MEIDNet targets: total magnetization per formula unit, Tc (log scale) as scalar targets</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally SuperCon (NIMS) and 3DSC experimental about 16,000 Tc entries; 3DSC links 5,700 to structures experimental superconducting critical temperatures; 3DSC matches them to ICSD / Materials Project structures 3DSC is ready (structure + Tc); use log Tc as the target MAGNDATA (Bilbao) experimental about 2,000 magnetic structures experimentally determined commensurate and incommensurate magnetic structures structures with magnetic ordering; join with moments as targets"},{"location":"explore/databases.html#mechanical-dielectric-and-piezoelectric","title":"Mechanical, dielectric and piezoelectric","text":"<p>What to look for: bulk and shear moduli, hardness, dielectric tensor, piezoelectric coefficients, phonon stability</p> <p>As MEIDNet targets: bulk modulus, dielectric constant or a piezoelectric scalar as targets</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Materials Cloud computed curated archives (MC3D, MC2D, phonons, ...) MC3D relaxed structures, MC2D exfoliable monolayers, phonon database, Sssp pseudopotential sets, workflow provenance (AiiDA) ready for MC3D / MC2D tables Matbench and Matbench Discovery computed 13 tasks (Matbench); 257,000 WBM test structures (Discovery) standardised property-prediction tasks (gap, formation energy, moduli, dielectric, ...) with fixed folds; Discovery ranks universal potentials on stability prediction with F1, DAF, RMSD and a leaderboard with plots use the tasks as property tables; the Discovery leaderboard is the model for MEIDNet Benchmarks matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#2d-materials","title":"2D materials","text":"<p>What to look for: exfoliation energy, gap, magnetic state, dynamic stability of monolayers</p> <p>As MEIDNet targets: a 2D prototype family (e.g. MXene or TMD) with gap and exfoliation energy as targets</p> database kind size features use with MEIDNet JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) Materials Cloud computed curated archives (MC3D, MC2D, phonons, ...) MC3D relaxed structures, MC2D exfoliable monolayers, phonon database, Sssp pseudopotential sets, workflow provenance (AiiDA) ready for MC3D / MC2D tables C2DB computed about 4,000 monolayers stability (dynamic and thermodynamic), PBE / HSE / GW gaps, magnetic state, optical absorption, piezoelectric and Raman data for monolayers ready for a 2D prototype family 2DMatPedia computed about 6,000 monolayers exfoliation energy, band gap, magnetic moment of monolayers derived from bulk databases ready for a 2D prototype family"},{"location":"explore/databases.html#porous-materials-and-mofs","title":"Porous materials and MOFs","text":"<p>What to look for: pore geometry, gas uptake, band gap of the framework, stability</p> <p>As MEIDNet targets: frameworks exceed MEIDNet's cell-size limit today (max_sites); use the property tables to set targets and the descriptors as inspiration</p> database kind size features use with MEIDNet Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files Cambridge Structural Database (CSD) experimental more than 1,300,000 organic and metal-organic structures experimental molecular and MOF crystal structures beyond today's cell-size limit; descriptors only QMOF computed about 20,000 MOFs DFT-optimised MOF structures with band gaps and charges beyond today's cell-size limit; targets and descriptors CoRE MOF experimental about 14,000 experimental MOFs computation-ready experimental MOF structures with pore descriptors beyond today's cell-size limit"},{"location":"explore/databases.html#generative-model-benchmarks","title":"Generative-model benchmarks","text":"<p>What to look for: fixed splits used by CDVAE, DiffCSP, MatterGen and MEIDNet; validity, uniqueness, novelty and stability metrics</p> <p>As MEIDNet targets: the Perov-5 split is the published MEIDNet benchmark; MP-20 and Carbon-24 test structure representation</p> database kind size features use with MEIDNet OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) GNoME (DeepMind) computed about 380,000 stable crystals structures predicted stable with DFT energies and hull distances; no electronic properties structures only: join with a computed property before training Cubic perovskites (CMR, Castelli et al.) computed about 19,000 ABX3 compositions formation (heat of formation) energies and direct / indirect gaps (GLLB-SC) for cubic ABX3 with O, N, S, F anions and their mixtures - the source of Perov-5 the published MEIDNet benchmark (Perov-5 split): meidnet download-data CDVAE splits (Perov-5, MP-20, Carbon-24) computed 18,928 / 45,231 / 10,153 structures fixed train / val / test splits used by generative-model papers; Perov-5 carries formation energy and band gap, the others energies only ready: the format meidnet reads directly (id, cif, property columns) Matbench and Matbench Discovery computed 13 tasks (Matbench); 257,000 WBM test structures (Discovery) standardised property-prediction tasks (gap, formation energy, moduli, dielectric, ...) with fixed folds; Discovery ranks universal potentials on stability prediction with F1, DAF, RMSD and a leaderboard with plots use the tasks as property tables; the Discovery leaderboard is the model for MEIDNet Benchmarks"},{"location":"explore/databases.html#all-databases","title":"All databases","text":"database kind access licence applications Materials Project computed web, mp-api (free key) CC BY 4.0 Solar cells and photovoltaics, Semiconductor physics, Batteries and ionic conductors, Magnetism and superconductivity, Mechanical, dielectric and piezoelectric, Catalysis and surfaces OQMD computed web, REST API, full download open Solar cells and photovoltaics, Semiconductor physics, Magnetism and superconductivity, Generative-model benchmarks AFLOW computed web, REST / AFLUX API open Semiconductor physics, Mechanical, dielectric and piezoelectric, Thermoelectrics, Magnetism and superconductivity JARVIS-DFT (NIST) computed web, jarvis-tools (figshare downloads) open (NIST) Solar cells and photovoltaics, Semiconductor physics, Thermoelectrics, Mechanical, dielectric and piezoelectric, 2D materials, Magnetism and superconductivity Alexandria computed bulk download CC BY 4.0 Generative-model benchmarks, Semiconductor physics, Magnetism and superconductivity, 2D materials GNoME (DeepMind) computed download CC BY-NC 4.0 Generative-model benchmarks NOMAD computed web, API, raw calculation files CC BY 4.0 Semiconductor physics, Solar cells and photovoltaics, Catalysis and surfaces Materials Cloud computed web, archive downloads CC BY 4.0 (per archive) 2D materials, Mechanical, dielectric and piezoelectric, Semiconductor physics C2DB computed web, ASE database download CC BY 4.0 2D materials, Semiconductor physics, Solar cells and photovoltaics 2DMatPedia computed web, download open 2D materials Cubic perovskites (CMR, Castelli et al.) computed web, ASE database download CC BY 4.0 Solar cells and photovoltaics, Semiconductor physics, Generative-model benchmarks CDVAE splits (Perov-5, MP-20, Carbon-24) computed download (CSV with CIF column) MIT (repository) Generative-model benchmarks Matbench and Matbench Discovery computed web, matbench / matbench-discovery packages open Generative-model benchmarks, Semiconductor physics, Mechanical, dielectric and piezoelectric matminer datasets computed + experimental matminer.datasets.load_dataset per dataset Semiconductor physics, Solar cells and photovoltaics, Thermoelectrics, Mechanical, dielectric and piezoelectric Crystallography Open Database (COD) experimental web, full download, REST public domain (CC0) Semiconductor physics, Solar cells and photovoltaics, Batteries and ionic conductors, Magnetism and superconductivity, Mechanical, dielectric and piezoelectric, Porous materials and MOFs ICSD experimental licence (most universities have it) commercial Semiconductor physics, Solar cells and photovoltaics, Batteries and ionic conductors, Magnetism and superconductivity, Mechanical, dielectric and piezoelectric Cambridge Structural Database (CSD) experimental licence commercial (CSD-Community subset free) Porous materials and MOFs Perovskite Database Project experimental web, download CC BY 4.0 Solar cells and photovoltaics Hybrid organic-inorganic perovskites (Kim et al. 2017) computed download (Scientific Data), matminer CC BY 4.0 Solar cells and photovoltaics, Semiconductor physics Experimental band gaps (Zhuo et al. 2018) experimental paper SI, matminer expt_gap per paper Semiconductor physics, Solar cells and photovoltaics Liverpool Ionic Conductivity Database (LiIonDB) experimental web, download open Batteries and ionic conductors Materials Project battery explorer computed web, mp-api CC BY 4.0 Batteries and ionic conductors UCSB thermoelectrics (Gaultois et al.) experimental matminer ucsb_thermoelectrics open Thermoelectrics Starrydata experimental web, download open Thermoelectrics Open Catalyst (OC20 / OC22) computed download, fairchem CC BY 4.0 Catalysis and surfaces Catalysis-Hub computed web, GraphQL API open Catalysis and surfaces SuperCon (NIMS) and 3DSC experimental NIMS MDR, 3DSC on GitHub open (3DSC) Magnetism and superconductivity MAGNDATA (Bilbao) experimental web open Magnetism and superconductivity QMOF computed download (figshare) CC BY 4.0 Porous materials and MOFs, Semiconductor physics CoRE MOF experimental download CC BY 4.0 Porous materials and MOFs Materials Data Facility computed + experimental web, API per dataset Semiconductor physics, Solar cells and photovoltaics, Batteries and ionic conductors, Thermoelectrics, Catalysis and surfaces, Mechanical, dielectric and piezoelectric <p>Missing one you use? Open an issue or edit <code>benchmarks/catalog/databases.json</code>.</p>"},{"location":"explore/datasets.html","title":"Datasets","text":"<p>Datasets that work with MEIDNet, and what has been done with each. One card per dataset, never one score across datasets: the tasks, sizes and properties differ, so results are only compared within a dataset (see Benchmarks).</p> dataset structures properties used size with MEIDNet Perov-5 cubic ABX\u2083 perovskites, 5 atoms per cell direct band gap, formation enthalpy 18,928 trained model, configuration, tutorial, benchmark MP-20 Materials Project, up to 20 atoms none aligned in the paper about 45,000 structure representation, published Carbon-24 carbon allotropes, up to 24 atoms none aligned in the paper about 10,000 structure representation, published Double perovskites A\u2082BB\u2032X\u2086 a family file, no dataset none 10,800 compositions (halide variant) design space and rules only Your dataset one material family any scalar columns any supported <p>New to multimodal data? Learn multimodality explains what a modality is, and Databases by application lists 31 computed and experimental sources.</p> <ul> <li> <p>Perov-5 \u2014 published multimodal benchmark</p> <p>18,928 cubic ABX\u2083 perovskites (CDVAE split: 11,356 train / 3,787 validation / 3,785 test), each with a relaxed structure, formation enthalpy (<code>heat_all</code>, eV/atom) and direct band gap (<code>dir_gap</code>, eV).</p> <p>Modalities used in MEIDNet: crystal structure \u00b7 electronic property (band gap) \u00b7 thermodynamic property (formation enthalpy).</p> <p>What the paper does with it: multimodal alignment of the structure and property encoders, property reconstruction, and the inverse-design demonstration (candidates generated from property targets, screened with MACE and validated by DFT).</p> <p>In this repository: <code>meidnet download-data</code> fetches it; the published checkpoint <code>checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth</code> was trained on it; it is the data behind the live Studio.</p> <p>Suitable tasks: inverse design from a band gap and a stability target \u00b7 property prediction from the structure \u00b7 structure\u2013property retrieval \u00b7 measuring how well two modalities align.</p> <p>MEIDNet implementation: \u2713 pretrained model \u00b7 \u2713 configuration (<code>examples/perov5/meidnet.yaml</code>) \u00b7 \u2713 tutorial (the paper's experiment, Colab) \u00b7 \u2713 benchmark results.</p> <p>Open in the Studio</p> <p>Benchmark results \u00b7 The paper's experiment \u00b7 Dataset source (CDVAE) \u00b7 Castelli et al. 2012, Xie et al. 2022</p> </li> <li> <p>MP-20 \u2014 published: structure-representation generalization</p> <p>~45,000 structures from the Materials Project with at most 20 atoms per cell (CDVAE split).</p> <p>Published MEIDNet use: generalization of the crystal encoder / decoder beyond perovskites (structure representation and reconstruction). It is not a multimodal benchmark in the paper: no property modality was aligned on it, and no MP-20 checkpoint is shipped here.</p> <p>Benchmark results \u00b7 Dataset source (CDVAE)</p> </li> <li> <p>Carbon-24 \u2014 published: structure-representation generalization</p> <p>~10,000 carbon structures with up to 24 atoms per cell (CDVAE split).</p> <p>Published MEIDNet use: the same structure-representation test as MP-20. Not a multimodal benchmark; no Carbon-24 checkpoint is shipped here.</p> <p>Benchmark results \u00b7 Dataset source (CDVAE)</p> </li> </ul>"},{"location":"explore/datasets.html#your-dataset","title":"Your dataset","text":"<p>Use your dataset in the Studio</p> <p>Any table with an id, one or more scalar property columns and a crystal structure per row (CIF text in a column or one <code>.cif</code> file per id) can be used directly: upload it in the Studio or run <code>meidnet init</code> \u2192 <code>check</code> \u2192 <code>train</code> \u2192 <code>generate</code> (bring your own dataset). The structures must belong to one material family (prototype + site groups), which is what the rules and the search need.</p>"},{"location":"explore/datasets.html#datasets-we-would-like-to-see-benchmarked","title":"Datasets we would like to see benchmarked","text":"<p>Battery conductors, MOFs, 2D materials and spinels are natural next families. If you run MEIDNet on one of them, contribute the result: the dataset gets its own benchmark page, and a verified row once the result has been reproduced here.</p>"},{"location":"learn/index.html","title":"Learn multimodality","text":"<p>A material can be described in many ways: by its atoms and their positions, by a formula, by measured or computed properties, by a diffraction pattern, a spectrum, a micrograph or a paragraph in a paper. Each way of describing it is a modality. Multimodal learning trains one model on several modalities of the same materials, so that it can connect them: read a structure and predict a property, or \u2014 the point of MEIDNet \u2014 start from the properties you want and propose a structure.</p> <p>This short course takes about ten minutes. It explains the ideas behind MEIDNet and the alternatives to it, and every item says what works in MEIDNet today.</p> <ol> <li>What is a modality?</li> <li>The multimodality map (interactive)</li> <li>Five challenges of multimodal learning</li> <li>Contrastive learning: how two modalities are aligned</li> <li>From a shared space to inverse design</li> <li>Where to go next</li> </ol>"},{"location":"learn/index.html#what-is-a-modality","title":"1. What is a modality?","text":"<p>A modality is one kind of observation of a material, with its own shape of data. The shape decides which kind of neural network can read it.</p> modality what the data looks like example a network that reads it Crystal structure atoms with positions in a unit cell: a graph a CIF file graph neural network Composition elements and their fractions: a set CsPbI\u2083 set or element network Scalar properties a short vector of numbers band gap, formation enthalpy small fully connected network (MLP) Diffraction pattern a curve: intensity versus angle 2\u03b8 powder XRD 1D convolutional network Spectra and DOS a curve: signal versus energy density of states, XAS, Raman 1D convolutional network or transformer Image a grid of pixels SEM or TEM micrograph 2D convolutional network or vision transformer Text a sequence of words a synthesis procedure language model <p>Why combine them? Each modality sees part of the material. A structure fixes the atoms but does not tell you the band gap without a calculation; a band gap says nothing about which atoms produce it. A model that learns both together can translate between them in either direction.</p> <p>What it needs: paired data, where several modalities describe the same material. In Perov-5, every one of the 18,928 structures comes with its DFT band gap and formation enthalpy, which makes it a two-modality dataset.</p>"},{"location":"learn/index.html#the-multimodality-map","title":"2. The multimodality map","text":"<p>Click a modality to see what it is, which networks read it, how it is combined with others and its status in MEIDNet. Click the centre for the idea that ties them together.</p> The map as a table modality status in MEIDNet how Crystal structure Supported an E(n)-equivariant graph network (EGNN) encodes up to <code>max_sites</code> atoms (default 20) Scalar properties Supported all numeric columns form one vector, read by one MLP; every column becomes a target Composition Inside the structure the element on each site is part of the structure; formula-only data needs a composition model Processing conditions Partly numeric conditions can be extra scalar columns, encoded together with the properties Diffraction (XRD) Planned a vector-modality encoder for binned patterns, first item of the roadmap Spectra and DOS Planned as vector modalities Microscopy images Planned an image encoder into the shared space Text Planned a text encoder into the shared space <p>The same statuses, generated from the code, are on the capabilities page.</p>"},{"location":"learn/index.html#five-challenges","title":"3. Five challenges of multimodal learning","text":"<p>Baltru\u0161aitis, Ahuja and Morency organise multimodal machine learning around five challenges. Every multimodal model answers each of them in some way, even if only by leaving it out.</p> challenge the question a materials example the answer in MEIDNet Representation How is each modality encoded, and do they share one space (joint) or keep separate spaces tied by a constraint (coordinated)? a graph network for the structure, an MLP for the property vector both: two encoders whose outputs are coordinated by contrastive learning and then averaged into one joint latent of 128 numbers Translation How is one modality mapped to another? structure \u2192 properties (prediction), properties \u2192 structure (inverse design) both directions: decoders read the joint latent and the property latent alone Alignment Which parts of one modality correspond to which parts of another? which XRD peaks come from which lattice planes only whole materials: each structure is matched with its own property vector, not part by part Fusion How are the modalities joined to make a prediction? one descriptor from structure and band gap together the structure latent and the property latent are averaged (a note on the word \"early\") Co-learning Can one modality help learn another, for example a scarce one? learn from many simulated XRD patterns to read a few measured ones closest analogue: the property branch is trained to rebuild the crystal on its own, so at design time properties alone are enough <p>Further reading: Baltru\u0161aitis et al. 2019 (the taxonomy) \u00b7 Guo et al. 2019 (deep multimodal representation learning) \u00b7 Ngiam et al. 2011 (shared representations between modalities).</p>"},{"location":"learn/index.html#contrastive-learning","title":"4. Contrastive learning: how two modalities are aligned","text":"<p>Take a batch of \\(N\\) materials. Each gives two vectors of length one: \\(z_c\\) from its structure and \\(z_p\\) from its properties. The dot product of two such vectors is their cosine similarity. Lay out all \\(N \\times N\\) similarities in a table: the diagonal holds the true pairs, everything else is a structure next to someone else's properties.</p> <p>Contrastive learning turns each row into a multiple-choice question: which of the \\(N\\) property vectors belongs to this structure? The right answer is on the diagonal. It asks the same question for each column, from properties to structures. The loss MEIDNet uses is this symmetric InfoNCE loss:</p> \\[ \\mathcal{L}_\\text{InfoNCE} = \\frac{1}{2N}\\sum_{i=1}^{N}\\left[ -\\log\\frac{\\exp\\!\\big(z_c^{(i)}\\!\\cdot z_p^{(i)}/\\tau\\big)}{\\sum_{j=1}^{N}\\exp\\!\\big(z_c^{(i)}\\!\\cdot z_p^{(j)}/\\tau\\big)} -\\log\\frac{\\exp\\!\\big(z_p^{(i)}\\!\\cdot z_c^{(i)}/\\tau\\big)}{\\sum_{j=1}^{N}\\exp\\!\\big(z_p^{(i)}\\!\\cdot z_c^{(j)}/\\tau\\big)} \\right] \\] <p>The temperature \\(\\tau\\) sets how sharply the model must prefer the right answer. Move the slider to see its effect on a batch of five materials: each row shows how much probability the model gives to each candidate.</p> setting in MEIDNet value in <code>meidnet.yaml</code> temperature \\(\\tau\\) 0.01 <code>training.temperature</code> strength of the alignment 5.0 <code>training.contrastive_weight</code> warm-up: the strength grows from 0 to its full value over 60 % of the epochs (the paper: 1,200 epochs) <code>training.contrastive_warmup_epochs</code> size of the shared space 128 <code>model.latent_dim</code> <p>The warm-up is the paper's curriculum: the decoders first learn to rebuild crystals and properties, then the two latents are pulled together. After training, the training report measures the alignment with retrieval: how often the nearest property vector to a structure is its own (top 1) or among the five nearest (top 5). The Studio's Model block plots the mean cosine between matched pairs during training.</p> <p>The same loss pairs pictures with their captions in CLIP; MEIDNet pairs crystals with their properties.</p> <p>Further reading: van den Oord et al. 2018 (InfoNCE) \u00b7 Radford et al. 2021 (CLIP) \u00b7 Contrastive learning in the Architecture Atlas.</p>"},{"location":"learn/index.html#inverse-design","title":"5. From a shared space to inverse design","text":"<p>Once both modalities live in one space, a point in it can be read in both directions. MEIDNet makes the reverse direction possible with one extra training term: the crystal decoder is also trained from the property latent alone. A vector of wanted properties can then be encoded and decoded into a structure.</p> <p>The search in practice:</p> <ol> <li>Encode the target. The property encoder turns the target values into a starting point in the shared space.</li> <li>Optimise a population of latent points so that their decoded properties approach the target, while staying    diverse and near the target's latent.</li> <li>Decode one element per site of a material family (a prototype such as cubic ABX\u2083), so candidates are    compositions placed on a known arrangement of sites.</li> <li>Check the rules (charge balance, tolerance factor, distances and your own) and keep only what passes.</li> <li>Rank by distance to the target and save each candidate as a CIF with a report.</li> </ol> <p>The full objective and every step are in How MEIDNet works. Two limits matter: MEIDNet does not invent new atomic arrangements, and its predicted properties are estimates to confirm with DFT or experiment (scope and roadmap).</p> <p>Further reading: the MEIDNet paper (Babu et al. 2026) \u00b7 Moro, Loh et al. 2025 (multimodal foundation models for materials).</p>"},{"location":"learn/index.html#where-next","title":"6. Where to go next","text":"<ul> <li> <p>Architecture Atlas</p> <p>Early fusion, late fusion, shared latent spaces, cross-attention and contrastive learning, drawn the same way, with an advisor that recommends one for your data.</p> <p> Architecture Atlas</p> </li> <li> <p>Recipes by problem</p> <p>From a research question to an architecture, a dataset, a configuration and a tutorial.</p> <p> Recipes by problem</p> </li> <li> <p>Build it</p> <p>Upload a table of structures and properties, train in the browser, set targets, run the search.</p> <p> Open the Studio</p> </li> <li> <p>Find data</p> <p>Datasets that work with MEIDNet, and 31 computed and experimental databases grouped by application.</p> <p> Datasets</p> </li> </ul>"},{"location":"learn/index.html#references","title":"References","text":"<ul> <li>T. Baltru\u0161aitis, C. Ahuja, L.-P. Morency, \"Multimodal machine learning: a survey and taxonomy\",   IEEE Trans. Pattern Anal. Mach. Intell. 41, 423\u2013443 (2019). doi:10.1109/TPAMI.2018.2798607</li> <li>W. Guo, J. Wang, S. Wang, \"Deep multimodal representation learning: a survey\", IEEE Access 7, 63373\u201363394   (2019). doi:10.1109/ACCESS.2019.2916887</li> <li>J. Ngiam et al., \"Multimodal deep learning\", ICML (2011).</li> <li>A. van den Oord, Y. Li, O. Vinyals, \"Representation learning with contrastive predictive coding\" (2018).   arXiv:1807.03748</li> <li>A. Radford et al., \"Learning transferable visual models from natural language supervision\" (CLIP), ICML   (2021). arXiv:2103.00020</li> <li>V. G. Satorras, E. Hoogeboom, M. Welling, \"E(n) equivariant graph neural networks\", ICML (2021).   arXiv:2102.09844</li> <li>V. Moro, C. Loh et al., \"Multimodal foundation models for material property prediction and discovery\",   Newton (2025). doi:10.1016/j.newton.2025.100016</li> <li>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, \"MEIDNet: Multimodal generative AI framework for   inverse materials design\", npj Comput. Mater. (2026). doi:10.1038/s41524-026-02153-3</li> </ul>"},{"location":"learn/architectures.html","title":"Architecture Atlas","text":"<p>Five ways to build a model from several modalities, drawn the same way: inputs on the left, what the model produces on the right. Each card says when the design fits, what it costs, an example from materials science and whether MEIDNet implements it. At the end, the advisor recommends one for your data and your goal.</p> <p>New to the topic? Start with Learn multimodality.</p> Modality A Modality B concatenateab one model prediction Modality A Modality B model A model B guess A guess B combine average or vote Structure Properties EGNN MLP shared space (128-d) align average decoders structure andproperties A tokens (atoms) B tokens (spectrum) querieskeys, values cross-attention fused tokens\u2192 prediction properties of the batch structures true pairs:pulled together other pairs:pushed apart"},{"location":"learn/architectures.html#early-fusion","title":"Early fusion","text":"<p>Not in MEIDNet</p> <p>Idea. Join the modalities at the input: turn each into features, concatenate them into one vector, train one model on it.</p> <p>When it fits. Every material has every modality; the features have a fixed length and similar scales; you want a quick baseline.</p> <p>Strengths. Simple. The model can use interactions between features from the first layer. Weaknesses. A missing modality breaks the input. Very different shapes (a graph and a curve) are hard to concatenate. A large modality can drown a small one.</p> <p>Materials example. Composition descriptors, the processing temperature and a few peak positions from an XRD pattern, concatenated and given to a gradient-boosted regressor.</p> <p>References: Snoek et al. 2005; Baltru\u0161aitis et al. 2019.</p>"},{"location":"learn/architectures.html#late-fusion","title":"Late fusion","text":"<p>Not in MEIDNet</p> <p>Idea. One model per modality, each makes its own prediction; the predictions are combined by averaging, voting or a small model on top.</p> <p>When it fits. Modalities are often missing, models for each already exist, or robustness matters more than squeezing out the last bit of accuracy.</p> <p>Strengths. Modular. Tolerates a missing modality. Each model can be trained on its own, larger dataset. Weaknesses. Cannot learn how the modalities interact. There is no shared representation, so no translation from one modality to another and no inverse design.</p> <p>Materials example. A structure-based graph network and a spectrum-based convolutional network each predict a phase label; the final label averages their probabilities.</p> <p>References: Snoek et al. 2005; Baltru\u0161aitis et al. 2019.</p>"},{"location":"learn/architectures.html#shared-latent","title":"Shared latent space (joint and coordinated)","text":"<p>Core idea used by MEIDNet Supported</p> <p>Idea. One encoder per modality maps into a common space. In a coordinated representation the latents stay separate but are trained to agree; in a joint representation they are merged into one vector that the decoders read. MEIDNet does both: the structure latent and the property latent are aligned by contrastive learning, then averaged into the joint latent; the decoders read the joint latent and the property latent alone.</p> <p>When it fits. You want to translate between modalities (properties \u2192 structure for inverse design), retrieve one modality from another, or search a space that both understand. You have paired examples.</p> <p>Strengths. Translation in both directions, retrieval, and one space to optimise in. Each modality keeps the encoder that suits its shape. Weaknesses. Needs paired data. The space is only as good as the alignment. Several losses must be balanced.</p> <p>Materials example. MEIDNet: crystal structures and DFT properties of 18,928 cubic perovskites, searched for compositions with a target band gap and formation enthalpy (the paper's experiment).</p> <p>References: Ngiam et al. 2011; Baltru\u0161aitis et al. 2019; Babu et al. 2026; Moro, Loh et al. 2025.</p>"},{"location":"learn/architectures.html#cross-attention","title":"Cross-attention and transformer fusion","text":"<p>Not in MEIDNet</p> <p>Idea. Each modality becomes a sequence of tokens (atoms, segments of a spectrum, words). Attention lets every token of one modality look up the tokens of the other that matter to it, layer after layer.</p> <p>When it fits. The modalities have parts that interact in detail and are not aligned in advance; the dataset is large.</p> <p>Strengths. Learns which parts relate to which, and handles sequences of different lengths without manual alignment. Weaknesses. Needs much data and compute; the cost grows with the product of the two sequence lengths; harder to interpret than a single shared vector.</p> <p>Materials example. Spectrum segments attend to the atoms of the structure, so the model learns which sites shape which spectral features.</p> <p>References: Tsai et al. 2019 (Multimodal Transformer); Lu et al. 2019 (ViLBERT).</p>"},{"location":"learn/architectures.html#contrastive","title":"Contrastive learning","text":"<p>Used by MEIDNet Supported</p> <p>Idea. A training objective rather than a wiring. In a batch of paired examples, each example's two embeddings must be more similar to each other than to any other example's. It is how the latents of a shared space are coordinated.</p> <p>When it fits. You have pairs (a structure and its properties, a pattern and its structure) and want a space in which nearest neighbours are meaningful across modalities.</p> <p>Strengths. Learns from the pairing alone; gives retrieval for free; scales to large datasets. Weaknesses. Needs batches with enough other examples; sensitive to the temperature; two genuinely similar materials in one batch are still pushed apart.</p> <p>In MEIDNet. Symmetric InfoNCE with temperature 0.01 and weight 5, switched on gradually over the warm-up (the paper's curriculum). The formula and a playground.</p> <p>References: van den Oord et al. 2018 (InfoNCE); Radford et al. 2021 (CLIP).</p>"},{"location":"learn/architectures.html#side-by-side","title":"Side by side","text":"early fusion late fusion shared latent cross-attention contrastive (objective) where the modalities meet input features predictions a common latent space inside the network, token by token the latent space, through the loss every modality needed for every sample yes no paired examples for training usually paired examples learns interactions between modalities yes no yes yes, in detail for whole samples translates one modality into another no no yes with a decoder retrieval data needed small to medium small per model medium large medium to large in MEIDNet no no yes, the core no yes <p>Beyond these five. Conditional generative models such as diffusion (MatterGen) or variational autoencoders (CDVAE) also translate from properties to structures, and they generate free atomic arrangements, which MEIDNet does not. MEIDNet instead decodes compositions onto a prototype family and checks them against chemical rules.</p>"},{"location":"learn/architectures.html#advisor","title":"Which architecture should I use?","text":"<p>Tick the data you have and choose what you want to do. The recommendation says why, what it costs and whether MEIDNet can do it today.</p> The advisor needs JavaScript. In short: structures with scalar properties, for design or retrieval \u2192 shared latent space with contrastive alignment (MEIDNet). Formula only \u2192 a composition model such as Roost. A pattern or spectrum to identify a phase \u2192 a 1D convolutional classifier. Modalities that are often missing \u2192 late fusion. Long sequences that interact in detail \u2192 cross-attention."},{"location":"learn/architectures.html#terminology","title":"A note on words","text":"<p>\"Early fusion\" in the MEIDNet paper names the model in which the structure latent and the property latent are averaged into one joint latent: <code>z_joint = (z_c + z_p) / 2</code> in <code>meidnet/model.py</code>. In the survey literature, early fusion usually means joining raw features at the input, as in the first card; averaging two learned latents, each from its own encoder, is intermediate or model-level fusion, the shared-latent card. Both names describe the same code. The checkpoints shipped here are the paper's early-fusion models, and the code implements only this fusion: there is no late-fusion option in MEIDNet.</p> <p>\"Curriculum\" in the paper is the contrastive warm-up: the alignment weight grows from zero to its full value over <code>training.contrastive_warmup_epochs</code>, so that the decoders learn first.</p>"},{"location":"learn/architectures.html#references","title":"References","text":"<ul> <li>C. G. M. Snoek, M. Worring, A. W. M. Smeulders, \"Early versus late fusion in semantic video analysis\",   ACM Multimedia (2005). doi:10.1145/1101149.1101236</li> <li>T. Baltru\u0161aitis, C. Ahuja, L.-P. Morency, \"Multimodal machine learning: a survey and taxonomy\",   IEEE TPAMI 41, 423\u2013443 (2019). doi:10.1109/TPAMI.2018.2798607</li> <li>J. Ngiam et al., \"Multimodal deep learning\", ICML (2011).</li> <li>Y.-H. H. Tsai et al., \"Multimodal Transformer for unaligned multimodal language sequences\", ACL (2019).   doi:10.18653/v1/P19-1656</li> <li>J. Lu, D. Batra, D. Parikh, S. Lee, \"ViLBERT: pretraining task-agnostic visiolinguistic representations for   vision-and-language tasks\", NeurIPS (2019). arXiv:1908.02265</li> <li>A. van den Oord, Y. Li, O. Vinyals, \"Representation learning with contrastive predictive coding\" (2018).   arXiv:1807.03748</li> <li>A. Radford et al., \"Learning transferable visual models from natural language supervision\", ICML (2021).   arXiv:2103.00020</li> <li>R. E. A. Goodall, A. A. Lee, \"Predicting materials properties without crystal structure: deep representation   learning from stoichiometry\" (Roost), Nat. Commun. 11, 6280 (2020). doi:10.1038/s41467-020-19964-7</li> <li>W. B. Park et al., \"Classification of crystal structure using a convolutional neural network\", IUCrJ 4,   486\u2013494 (2017). doi:10.1107/S205225251700714X</li> <li>C. Zeni et al., \"A generative model for inorganic materials design\" (MatterGen), Nature 639, 624\u2013632 (2025).   doi:10.1038/s41586-025-08628-5</li> <li>T. Xie et al., \"Crystal diffusion variational autoencoder for periodic material generation\" (CDVAE), ICLR   (2022). arXiv:2110.06197</li> <li>V. Moro, C. Loh et al., \"Multimodal foundation models for material property prediction and discovery\",   Newton (2025). doi:10.1016/j.newton.2025.100016</li> <li>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, \"MEIDNet: Multimodal generative AI framework for   inverse materials design\", npj Comput. Mater. (2026). doi:10.1038/s41524-026-02153-3</li> </ul>"},{"location":"learn/recipes.html","title":"Recipes by problem","text":"<p>Start from a research question. Each recipe goes problem \u2192 architecture \u2192 why \u2192 data \u2192 configuration \u2192 tutorial \u2192 references, and says plainly whether MEIDNet does it today. Where it does not, the recipe points to the kind of model that does.</p> <p>For step-by-step changes to a MEIDNet configuration (another target, another family, a new rule), see the how-to guides.</p> problem architecture MEIDNet today Perovskites with a target band gap that are likely to be stable shared latent + contrastive Supported Candidates from your own DFT results for one family shared latent + contrastive Supported Screen a whole family for a target before searching structure encoder of the shared space Supported Identify the crystal structure from an XRD pattern 1D convolutional network; contrastive pattern \u2194 structure Planned Predict a property when only the formula is known composition network Use another tool Connect synthesis text with structures text encoder + contrastive or cross-attention Planned Design new atomic arrangements conditional generative model over full geometry Use another tool"},{"location":"learn/recipes.html#target-band-gap","title":"Perovskites with a target band gap that are likely to be stable","text":"<p>Supported in MEIDNet</p> Problem Find cubic ABX\u2083 halide perovskites with a direct band gap near 1.5 eV and a negative formation enthalpy. Architecture Shared latent space with contrastive alignment: MEIDNet. Why The question runs from properties to structure. A shared space whose crystal decoder is also trained from the property latent alone gives that direction, and the family's rules keep the chemistry sensible. Data Perov-5: 18,928 structures with DFT band gap and formation enthalpy (dataset card). The published checkpoint is trained on it. Tutorial The paper's experiment \u00b7 in the browser: Studio, Targets block \u00b7 screen the candidates with MACE References Babu et al. 2026 (MEIDNet); Castelli et al. 2012 (the dataset). <p>One command with the published model:</p> <pre><code>meidnet demo --family halide --band-gap 1.5 --enthalpy -0.10\n</code></pre> <p>Or in <code>meidnet.yaml</code> (from <code>examples/perov5/meidnet.yaml</code>):</p> <pre><code>generation:\n  family: perovskite_abx3\n  variant: halide\n  objectives:\n    - {property: dir_gap,  loss: l2, weight: 10000, select_weight: 1.0}\n    - {property: heat_all, loss: l1, weight: 6000,  select_weight: 0.4}\n  targets:\n    - {dir_gap: 1.5, heat_all: -0.10}\n</code></pre>"},{"location":"learn/recipes.html#own-family","title":"Candidates from your own DFT results for one family","text":"<p>Supported in MEIDNet</p> Problem You computed structures and properties for one prototype family, for example A\u2082BB\u2032X\u2086 double perovskites, and want new compositions with chosen property values. Architecture Shared latent space with contrastive alignment, trained on your table. Why The same reasoning as above. Every scalar column you give becomes a target you can set, and the family file says which sites exist and which elements may sit on them. Data A table with an id, numeric property columns and one CIF per row. Public sources by application are on Databases by application. Tutorial Bring your own dataset \u00b7 in the browser: Studio, Data block \u00b7 change the material family References Babu et al. 2026. <pre><code>data:\n  table: my_results.csv\n  id_column: id\n  cif_column: cif\n  properties:\n    - {column: band_gap, unit: eV}\n    - {column: e_form,   unit: eV/atom}\ngeneration:\n  family: double_perovskite_a2bbx6\n  variant: halide\n  objectives:\n    - {property: band_gap, loss: l2,      weight: 10000}\n    - {property: e_form,   loss: at_most, weight: 6000}\n  targets:\n    - {band_gap: 1.8, e_form: -0.10}\n</code></pre> <p>Then <code>meidnet check</code> \u2192 <code>meidnet train</code> \u2192 <code>meidnet generate</code>; each step writes a report.</p>"},{"location":"learn/recipes.html#screen-family","title":"Screen a whole family for a target before searching","text":"<p>Supported in MEIDNet</p> Problem Before running a search, see every composition the family allows, its rule values and its predicted properties, and which are predicted closest to the target. Architecture The structure encoder of the shared space, read in the forward direction (structure \u2192 properties). Why For a prototype family the set of compositions is finite, so it can be scored exhaustively in seconds. Comparing these predictions with the search's candidates is a useful sanity check. Data A trained model and a family; nothing else. Tutorial Explore in 3D \u00b7 in the browser: Studio, design space \u00b7 How MEIDNet works References Babu et al. 2026. <pre><code>meidnet space examples/perov5/meidnet.yaml --model checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth -o space.csv\n</code></pre>"},{"location":"learn/recipes.html#xrd-structure","title":"Identify the crystal structure from an XRD pattern","text":"<p>Planned in MEIDNet</p> Problem Given a powder diffraction pattern, tell which crystal system, space group or known structure it comes from. Architecture Today: a 1D convolutional classifier on the pattern (one modality). Multimodal: a pattern encoder aligned with a structure encoder by contrastive learning, so a pattern retrieves the structures it matches. Why A pattern is a fingerprint of the lattice. A classifier needs labels; a contrastive model needs only (pattern, structure) pairs, which can be simulated from any structure database. Data Simulate patterns from computed or experimental structures (semiconductor physics, all databases), for example with pymatgen's XRD calculator; validate on measured patterns. In MEIDNet A vector-modality encoder for binned XRD is the first item of the roadmap. It is not implemented. References Park et al. 2017; Baltru\u0161aitis et al. 2019."},{"location":"learn/recipes.html#formula-only","title":"Predict a property when only the formula is known","text":"<p>Use another tool</p> Problem Predict a band gap or a stability measure for compositions whose crystal structure is unknown, as in many experimental tables. Architecture A composition network that reads the formula as a weighted set of elements, such as Roost. With extra numeric inputs (temperature, doping), early fusion of composition features and those numbers is a strong baseline. Why MEIDNet needs a structure for every row. If your formulas all sit on one prototype (for example ABX\u2083), you can build those structures on the prototype and use MEIDNet; otherwise a composition model is the right tool. Data Experimental band gaps and Matbench tasks are listed on Databases by application. References Goodall and Lee 2020 (Roost); Dunn et al. 2020 (Matbench)."},{"location":"learn/recipes.html#synthesis-text","title":"Connect synthesis text with structures","text":"<p>Planned in MEIDNet</p> Problem Link written synthesis procedures to the structures they produce, to suggest how a candidate might be made or which candidates resemble known syntheses. Architecture A text encoder (a language model) aligned with a structure encoder by contrastive learning; cross-attention when individual words must be tied to individual elements or steps. Why Text and structures have no common format; a shared space lets each be compared with the other. Data Text-mined synthesis recipes (Kononova et al. 2019) paired with structures from a structure database. In MEIDNet A text encoder into the shared space is planned, not implemented. References Kononova et al. 2019; Moro, Loh et al. 2025."},{"location":"learn/recipes.html#new-arrangements","title":"Design new atomic arrangements","text":"<p>Use another tool</p> Problem Generate crystals whose atomic arrangement is not one of a few known prototypes. Architecture A conditional generative model over the full geometry: diffusion (MatterGen) or a variational autoencoder (CDVAE). Why MEIDNet places compositions on a family prototype and does not invent arrangements (scope). Models that generate positions and cells directly can. Data Large structure databases; the generative benchmarks are listed under generative-model benchmarks. References Zeni et al. 2025 (MatterGen); Xie et al. 2022 (CDVAE)."},{"location":"learn/recipes.html#references","title":"References","text":"<ul> <li>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, \"MEIDNet: Multimodal generative AI framework for   inverse materials design\", npj Comput. Mater. (2026). doi:10.1038/s41524-026-02153-3</li> <li>I. E. Castelli et al., \"New cubic perovskites for one- and two-photon water splitting using the computational   materials repository\", Energy Environ. Sci. 5, 9034 (2012). doi:10.1039/C2EE22341D</li> <li>W. B. Park et al., \"Classification of crystal structure using a convolutional neural network\", IUCrJ 4,   486\u2013494 (2017). doi:10.1107/S205225251700714X</li> <li>R. E. A. Goodall, A. A. Lee, \"Predicting materials properties without crystal structure: deep representation   learning from stoichiometry\", Nat. Commun. 11, 6280 (2020). doi:10.1038/s41467-020-19964-7</li> <li>A. Dunn et al., \"Benchmarking materials property prediction methods: the Matbench   test set and Automatminer reference algorithm\", npj Comput. Mater. 6, 138 (2020). doi:10.1038/s41524-020-00406-3</li> <li>O. Kononova et al., \"Text-mined dataset of inorganic materials synthesis recipes\", Sci. Data 6, 203 (2019).   doi:10.1038/s41597-019-0224-1</li> <li>C. Zeni et al., \"A generative model for inorganic materials design\", Nature 639, 624\u2013632 (2025).   doi:10.1038/s41586-025-08628-5</li> <li>T. Xie et al., \"Crystal diffusion variational autoencoder for periodic material generation\", ICLR (2022).   arXiv:2110.06197</li> <li>T. Baltru\u0161aitis, C. Ahuja, L.-P. Morency, \"Multimodal machine learning: a survey and taxonomy\", IEEE TPAMI 41, 423\u2013443 (2019). doi:10.1109/TPAMI.2018.2798607</li> <li>V. Moro, C. Loh et al., \"Multimodal foundation models for material property prediction and discovery\",   Newton (2025). doi:10.1016/j.newton.2025.100016</li> </ul>"},{"location":"recipes/add-constraint.html","title":"Add a rule (constraint)","text":"<p>A rule is a function that looks at a candidate and returns pass/fail, the measured value, the allowed window and a sentence. The search enforces it; every report and the Studio explain it.</p>"},{"location":"recipes/add-constraint.html#1-write-it","title":"1. Write it","text":"<p><code>my_rules.py</code>:</p> <pre><code>from meidnet.constraints import CONSTRAINTS\n\n@CONSTRAINTS.register(\"no_lead\", \"Rejects any composition containing Pb.\")\ndef no_lead(cand):\n    return cand.result(\"no_lead\", \"Pb\" not in cand.elements.values())\n\n@CONSTRAINTS.register(\"mass_window\", \"Mean atomic mass must lie between min and max (g/mol).\")\ndef mass_window(cand, min=0.0, max=200.0):\n    from pymatgen.core import Element\n    m = sum(Element(e).atomic_mass for e in cand.structure.species) / len(cand.structure)\n    return cand.result(\"mass_window\", min &lt;= m &lt;= max, float(m), (min, max), f\"mean mass {m:.1f}\")\n</code></pre> <p>What a candidate gives you: <code>cand.elements</code> (<code>{group: element}</code>), <code>cand.structure</code> (pymatgen, after symmetry refinement), <code>cand.raw</code> (before), <code>cand.lattice_a</code>, <code>cand.predictions</code> (<code>{property: value}</code>), <code>cand.radius(group)</code>.</p>"},{"location":"recipes/add-constraint.html#2-register-the-file-and-name-the-rule","title":"2. Register the file and name the rule","text":"<pre><code>plugins: [my_rules.py]\ngeneration:\n  family: my_perovskite.yaml      # a copy of the built-in family with the rules appended:\n</code></pre> <pre><code>constraints:\n  - {name: charge_neutrality}\n  - ...\n  - {name: no_lead}\n  - {name: mass_window, min: 0, max: 120}\n</code></pre>"},{"location":"recipes/add-constraint.html#3-run","title":"3. Run","text":"<p><code>meidnet generate meidnet.yaml</code>. The candidate cards list No lead and mass window with the measured value; the funnel shows how many compositions each removed.</p> <p>A complete example lives in <code>examples/custom_rules/</code> (a no toxic elements rule and a max cell edge rule), and notebook 03 does the same in Colab.</p>"},{"location":"recipes/add-constraint.html#soft-terms","title":"Soft terms","text":"<p>Rules decide acceptance. To steer the latent search instead (a penalty with a gradient), register a search term with <code>@SEARCH_TERMS.register</code> in <code>meidnet.terms</code>; it receives a <code>SearchContext</code> with the decoder's outputs. See rules and search terms.</p>"},{"location":"recipes/add-property.html","title":"Add a property","text":"<p>MEIDNet's property modality is a vector of any length. Adding a property is a data change, not a code change.</p> <ol> <li>Add the column to your table.</li> <li> <p>List it under <code>data.properties</code> (label and unit are for the reports; <code>normalize: true</code> is the default and    recommended when properties have different scales):</p> <pre><code>data:\n  properties:\n    - {column: band_gap,     label: Band gap,          unit: eV}\n    - {column: formation_e,  label: Formation energy,  unit: eV/atom}\n    - {column: dielectric,   label: Dielectric constant}\n</code></pre> </li> <li> <p><code>meidnet check</code> \u2192 <code>meidnet train</code>. The training report rates each property separately.</p> </li> <li>Use it in <code>generation.objectives</code> and <code>targets</code>.</li> </ol> <p>Properties with many missing values: rows with a missing value are skipped (the check report counts them). Fill them, or drop the property.</p> <p>The published model</p> <p>The Perov-5 checkpoint has exactly two properties (formation enthalpy, direct band gap) and cannot steer others; train your own model to add some.</p>"},{"location":"recipes/change-family.html","title":"Change the material family","text":"<p>A family file tells MEIDNet what kind of crystal to build. The built-in ones are <code>meidnet/families/perovskite_abx3.yaml</code> and <code>double_perovskite_a2bbx6.yaml</code>; <code>meidnet families &lt;name&gt;</code> prints a summary. To use your own, copy one, edit it, and point to it:</p> <pre><code>generation:\n  family: my_family.yaml\n  variant: null          # or a variant you define\n</code></pre>"},{"location":"recipes/change-family.html#the-parts-of-a-family-file","title":"The parts of a family file","text":"<pre><code>name: spinel_ab2x4\ntitle: Normal spinel AB2X4 (Fd-3m, primitive cell)\n\nprototype:                    # WHERE atoms sit (fractional coordinates, in decoder slot order)\n  lattice: fcc_primitive      # cubic | fcc_primitive | bcc_primitive | explicit a,b,c,alpha,beta,gamma\n  reference_a: 8.1\n  sites:\n    - {group: A, frac: [0.125, 0.125, 0.125]}\n    - ...\n\ngroups:                       # WHO may sit there, and with which charges\n  A: {description: tetrahedral cation, elements: [Mg, Zn, Fe, Co, Ni, Mn], oxidation_states: {Mg: [2], ...}}\n  B: {...}\n  X: {...}\n\nsampling_order: [B, X, A]     # the LAST group is restricted to charge-balancing elements\n\nlattice:                      # HOW BIG the cell is\n  rule: bond_sum              # a = factor \u00d7 \u03a3 ionic radii of the listed groups (or rule: fixed)\n  groups: [B, X]\n  factor: 2.0\n  min: 3.0\n  max: 12.0\n\nvariants:                     # named sub-families: element subsets and rule parameters\n  oxide: {groups: {X: {elements: [O]}}, params: {tolerance_factor: {min: 0.8, max: 1.05}}}\n\nsearch_terms:                 # soft penalties during the latent search (see reference/rules)\n  - {name: site_separation, weight: 200}\n  - ...\n\nconstraints:                  # hard rules every saved candidate must pass\n  - {name: charge_neutrality}\n  - {name: min_distance, min: 0.8}\n  - ...\n</code></pre>"},{"location":"recipes/change-family.html#alignment-of-training-data","title":"Alignment of training data","text":"<p>For generation to mean anything, the model must have been trained on structures whose atoms are in the prototype's site order. <code>data.align_to_prototype: true</code> (default) re-orders every training structure to match the family named in <code>generation.family</code> (or top-level <code>family:</code>); structures that do not fit are reported by <code>meidnet check</code>. Raise <code>data.prototype_tolerance</code> for distorted cells.</p>"},{"location":"recipes/change-family.html#checking-a-family-file","title":"Checking a family file","text":"<pre><code>meidnet families my_family.yaml --variant oxide\nmeidnet space meidnet.yaml      # every composition with rule values \u2192 CSV\n</code></pre> <p>Errors name the problem: unknown element symbols, a prototype group without a definition, an override that matches no rule, a group left empty after your filters.</p>"},{"location":"recipes/change-family.html#worked-example","title":"Worked example","text":"<p>The double perovskite page walks through a 10-site, non-cubic-primitive family written this way.</p>"},{"location":"recipes/change-targets.html","title":"Change the target properties","text":""},{"location":"recipes/change-targets.html#a-different-value","title":"A different value","text":"<pre><code>generation:\n  targets:\n    - {band_gap: 1.2}\n    - {band_gap: 1.8}      # each entry is its own search\n</code></pre>"},{"location":"recipes/change-targets.html#a-different-kind-of-goal","title":"A different kind of goal","text":"<p><code>loss</code> says how a prediction is compared with the target:</p> <code>loss</code> meaning typical use <code>l2</code> squared distance (default) hit a value <code>l1</code> absolute distance hit a value, less sensitive to outliers <code>at_most</code> only values above the target are penalised formation enthalpy \u2264 \u22120.1 eV/atom <code>at_least</code> only values below are penalised dielectric constant \u2265 20 <pre><code>objectives:\n  - {property: band_gap,   loss: l2,       weight: 10000, select_weight: 1.0}\n  - {property: enthalpy,   loss: at_most,  weight: 6000,  select_weight: 0.4}\n</code></pre> <p><code>weight</code> acts during the latent search; <code>select_weight</code> when ranking finished candidates.</p>"},{"location":"recipes/change-targets.html#a-different-property","title":"A different property","text":"<p>The model knows the properties it was trained on (<code>meidnet info model.pt</code> lists them). To steer a property that is not in the model, add it to the data and retrain.</p>"},{"location":"recipes/change-targets.html#in-the-studio","title":"In the Studio","text":"<p>Drag the target slider; the predicted-closest table and the scatter update at once, and a warning appears if the target lies outside the training range. Export the YAML when you are happy.</p>"},{"location":"recipes/elements.html","title":"Exclude or restrict elements","text":"<pre><code>generation:\n  exclude_elements: [Pb, Cd, Hg]           # never used, on any site\n  only_elements:                           # restrict a site group to a shortlist\n    B: [Ti, Zr, Hf, Sn]\n</code></pre> <p>Both are applied to the family before the search starts, so the decoder cannot even propose an excluded element. The Studio's element chips do the same thing (click to toggle) and show the counts change.</p> <p>To change which elements a family allows at all, or their oxidation states, copy the family file and edit the <code>groups:</code> section \u2014 see Change the material family.</p>"},{"location":"recipes/screen.html","title":"Screen stability with MACE","text":"<pre><code>pip install \"meidnet[stability]\"            # ase + mace-torch (~500 MB of weights on first use)\nmeidnet screen runs/my_oxides/generation --train-csv materials.csv\n</code></pre> <p>For each CIF: relax cell and positions with the MACE-MP-0 universal potential, compute the formation energy per atom against elemental reference phases, and report</p> <ul> <li>S stable \u2014 \u0394H_f \u2264 0.10 eV/atom (Materials Project's \"potentially synthesisable\" criterion; <code>--threshold</code>),</li> <li>U unique \u2014 no structural duplicate among the candidates,</li> <li>N novel \u2014 reduced formula absent from the training table.</li> </ul> <p>Results go to <code>stability.csv</code> in the folder. Structures with |\u0394H_f| &gt; 5 eV/atom are flagged as artefacts and excluded from the rates. The reference phases and GGA-style O/N corrections are those of the paper.</p> <p>This is a first filter, not a verdict: MACE formation energies carry ~0.1 eV/atom uncertainty, and a cubic prototype may relax to a lower-symmetry structure.</p>"},{"location":"reference/api.html","title":"Python API","text":"<p>The command line is a thin layer over these functions; use them from notebooks or your own scripts.</p> <pre><code>from meidnet.config import load_config, config_from_dict\nfrom meidnet.pipeline import check, train, generate\n\ncfg = load_config(\"meidnet.yaml\")\ninfo = check(cfg)                 # dict: report (counts, skip reasons), records, family, report_path\ninfo = check(cfg, keep_structures=True)   # records also keep their pymatgen Structure (record.structure)\nmodel_path = train(cfg)           # writes model.pt + training_report.html, returns the path\nresult = generate(cfg)            # GenerationResult: .saved (candidates), .targets (funnel logs)\n</code></pre>"},{"location":"reference/api.html#pieces","title":"Pieces","text":"<pre><code>from meidnet.checkpoint import load_checkpoint, describe, property_ranges\nlm = load_checkpoint(\"model.pt\")          # LoadedModel: .model, .stats (property names/units/normalisation)\nlm.model.encode_crystal(x)                # structure \u2192 latent (x from meidnet.data.featurize)\nlm.model.encode_properties(p)             # normalised property vector \u2192 latent\n\nfrom meidnet.family import load_family\nfam = load_family(\"perovskite_abx3\", variant=\"halide\", exclude=[\"Pb\"], overrides={\"tolerance_factor\": {\"max\": 1.0}})\n\nfrom meidnet.constraints import build_candidate, evaluate, CONSTRAINTS\ncand = evaluate(build_candidate(fam, {\"A\": \"Cs\", \"B\": \"Sn\", \"X\": \"Br\"}), fam.constraints)\ncand.passed, [r.to_dict() for r in cand.results]\n\nfrom meidnet.generate import Designer\nfrom meidnet.config import GenerationSection\ng = GenerationSection(family=\"perovskite_abx3\", variant=\"halide\",\n                      objectives=[{\"property\": \"dir_gap\"}], targets=[{\"dir_gap\": 2.0}], rounds=3, steps=200)\nres = Designer(lm, fam, g, on_saved=lambda c, t: print(c.formula)).run(\"out/\")\n\nfrom meidnet.designspace import enumerate_space, space_to_csv\nspace = enumerate_space(fam, lm)          # every composition: rule values + predictions\n</code></pre> <p>Custom rules and search terms: <code>CONSTRAINTS.register</code>, <code>meidnet.terms.SEARCH_TERMS.register</code>.</p>"},{"location":"reference/api.html#extra-rules-in-the-configuration","title":"Extra rules in the configuration","text":"<p><code>generation.extra_constraints</code> adds rules to the family's own for one run, without editing the family file \u2014 this is how the Studio saves a predicted-property window:</p> <pre><code>generation:\n  family: perovskite_abx3\n  variant: halide\n  extra_constraints:\n    - {name: property_window, property: dir_gap, min: 1.5, max: 3.0}\n</code></pre> <p><code>meidnet.pipeline.family_for(cfg, need_variant=True)</code> returns the family a run uses, with these rules appended; <code>generation.family</code> / <code>generation.variant</code> take precedence over the top-level <code>family</code>.</p>"},{"location":"reference/api.html#3d-datasets-chemiscope","title":"3D datasets (chemiscope)","text":"<p><code>meidnet.studio.chemiscope</code> turns MEIDNet objects into chemiscope datasets \u2014 plain dicts you can save with <code>json.dump</code> and open at chemiscope.org, or show in a notebook with <code>chemiscope.show(...)</code>:</p> <pre><code>from meidnet.studio.chemiscope import (space_dataset, candidates_dataset, records_dataset,\n                                       published_data_dataset, structure_to_chemiscope)\n\nds = space_dataset(fam, space, lm, max_structures=2000)   # design space: formula, site_*, pred_*, rule_*, passes_all\nds = candidates_dataset(candidate_dicts, fam, run_dir, lm=lm)   # candidates of a run (reads their CIFs when present;\n                                                                # lm adds units and property names)\nds = records_dataset(info[\"records\"], cfg.data.properties, source=\"my data\")   # needs check(cfg, keep_structures=True)\nstructure_to_chemiscope(structure)                         # one pymatgen Structure \u2192 {size, names, x, y, z, cell}\n</code></pre> <p><code>candidate_dicts</code> are saved candidates as dicts \u2014 <code>[c.to_dict() for c in res.saved]</code> after a <code>Designer</code> run, or <code>[c for t in json.load(open(\"generation.json\"))[\"targets\"] for c in t[\"saved\"]]</code> from a finished run. <code>run_dir</code> is the run folder that contains <code>generation/</code> (<code>cfg.out_dir</code> after <code>meidnet generate</code>); where a CIF file is not found, the structure is rebuilt on the prototype from the candidate's elements.</p>"},{"location":"reference/api.html#scripting-the-studio","title":"Scripting the Studio","text":"<p>The Studio server is a small class you can drive without a browser \u2014 the 04 Colab notebook does exactly this. Every method takes the same JSON-like dicts the page sends; <code>session</code> keeps users (or experiments) apart.</p> <pre><code>import base64\nfrom meidnet.studio.server import Studio\n\nst = Studio(None)                                   # the published model; or Studio(load_config(\"meidnet.yaml\"))\nsid = \"my-session\"\nst.state(sid)                                       # model, ranges, families, limits, session info\n\nb64 = base64.b64encode(open(\"materials.csv\", \"rb\").read()).decode()\nup = st.upload({\"session\": sid, \"files\": [{\"name\": \"materials.csv\", \"b64\": b64}]})\nup[\"suggest\"]                                       # suggested id / CIF / property columns\n\nst.check_data({\"session\": sid, \"id_column\": \"material_id\", \"cif_column\": \"cif\",\n               \"properties\": [{\"column\": \"dir_gap\", \"unit\": \"eV\"}],\n               \"family\": \"perovskite_abx3\", \"variant\": \"halide\"})\nst.start_train({\"session\": sid, \"epochs\": 10, \"batch_size\": 16})\nst.status(sid, \"train\")                             # poll: running, progress (epoch, loss, curve), done, error\nst.select_model(sid, \"published\")                   # or \"own\"\n\nst.family({\"session\": sid, \"family\": \"perovskite_abx3\", \"variant\": \"halide\"})   # design space, scored\nst.start_search({\"session\": sid, \"generation\": {\"per_target\": 2, \"rounds\": 2, \"steps\": 100}})\nst.status(sid, \"search\")                            # candidates as they are found\nst.chemiscope(sid, \"space\", \"perovskite_abx3\", \"halide\")   # or \"candidates\" / \"data\"\nst.validate({\"yaml\": \"generation:\\n  rounds: 2\\n\"})        # {ok, generation, notes} or {ok: False, errors}\nst.export({\"session\": sid, \"generation\": {}})              # the meidnet.yaml text\n</code></pre>"},{"location":"reference/cli.html","title":"Command line","text":"<pre><code>meidnet init            write a starter meidnet.yaml next to your data\nmeidnet check  CONFIG   is my data usable?            \u2192 check_report.html\nmeidnet train  CONFIG   learn the latent space        \u2192 model.pt, training_report.html\nmeidnet generate CONFIG design candidates             \u2192 CIFs, generation_report.html\nmeidnet studio [CONFIG] interactive workbench in your browser\nmeidnet space  CONFIG   every composition of the family with rule values and predictions \u2192 CSV\nmeidnet demo            quick demo with the published perovskite model\nmeidnet screen DIR      stability screening of CIFs with MACE (optional extra)\nmeidnet info MODEL      describe a checkpoint\nmeidnet families [NAME] list / describe material families\nmeidnet schema          JSON Schema of the config (for editors and the Studio)\nmeidnet download-data   fetch the Perov-5 dataset used in the paper\n</code></pre> <p><code>meidnet &lt;command&gt; --help</code> lists the options. Useful ones:</p> <code>init --template perov5</code> the paper's configuration instead of a blank one <code>train --epochs N</code> override <code>training.epochs</code> <code>generate --quick</code> few rounds and steps \u2014 a smoke test <code>generate --model PATH</code> use another checkpoint <code>studio --export-static FILE.html</code> self-contained Studio for a website <code>studio --port 8765 --no-open</code> choose the port; do not open a browser <code>studio --public --host 0.0.0.0</code> host the Studio for other people: private session per visitor, capped budgets <code>studio --run-root DIR --docs-dir site</code> where sessions and searches are written; serve a built docs site at <code>/docs/</code> <code>space --chemiscope FILE.json</code> also write the design space as a 3D dataset for chemiscope.org <code>demo --family oxide --band-gap 3 --enthalpy -0.2 -n 5</code> demo targets <p>If <code>meidnet</code> is not on your PATH, <code>python -m meidnet.cli \u2026</code> is equivalent.</p>"},{"location":"reference/config.html","title":"Configuration reference","text":"<p>Every setting of <code>meidnet.yaml</code>, generated from the code (<code>meidnet schema</code> gives the JSON Schema). Unknown keys are rejected with a message naming the key, so typos cannot silently change a run.</p>"},{"location":"reference/config.html#top-level","title":"Top level","text":"setting type default meaning <code>name</code> string <code>meidnet_run</code> Name of this project; outputs go to output_dir. <code>description</code> string Free text shown at the top of reports. <code>output_dir</code> string <code>runs/{name}</code> Where checkpoints, CIFs and reports are written. <code>family</code> string/null <code>None</code> Family used to align training structures (defaults to generation.family). <code>model_path</code> string/null <code>None</code> Existing checkpoint to generate from (skip training). Default: /model.pt <code>plugins</code> array Python files that register your own constraints or search terms (see 'Add a constraint'). <code>data</code> object/null <code>None</code> <code>model</code> section <code>training</code> section <code>generation</code> object/null <code>None</code>"},{"location":"reference/config.html#data","title":"<code>data:</code>","text":"setting type default meaning <code>table</code> string required Path to a CSV/Excel/JSON table with one row per material. <code>id_column</code> string <code>material_id</code> Column with a unique identifier for each material. <code>cif_column</code> string/null <code>cif</code> Column that contains the CIF text of each structure. Set to null if you use structures_dir. <code>structures_dir</code> string/null <code>None</code> Folder with one CIF file per material, named .cif (used when there is no cif_column). <code>properties</code> array required Numeric columns that form the property modality (any number, at least one). <code>val_table</code> string/null <code>None</code> Optional separate validation table with the same columns. <code>val_fraction</code> number <code>0.1</code> Fraction of rows held out for validation when no val_table is given (0 = no validation). <code>max_sites</code> integer <code>20</code> Largest number of atoms per cell. Bigger structures are skipped. <code>neighbor_cutoff</code> number <code>4.0</code> Distance (\u00c5) below which two atoms count as neighbours. <code>align_to_prototype</code> boolean <code>True</code> Re-order the atoms of every structure so they match the family prototype's site order (needed for meaningful generation). The published Perov-5 model was trained without it. <code>prototype_tolerance</code> number <code>0.15</code> How far (fraction of a cell edge) an atom may sit from its prototype site and still count as that site. Raise it to accept distorted structures; lower it to keep only ideal ones."},{"location":"reference/config.html#dataproperties","title":"<code>data.properties[]</code>","text":"setting type default meaning <code>column</code> string required Name of the column in your table that holds this property. <code>label</code> string/null <code>None</code> Human-readable name used in reports (defaults to the column name). <code>unit</code> string Unit shown in reports, e.g. 'eV' or 'eV/atom'. <code>normalize</code> boolean <code>True</code> Standardise the property to zero mean and unit spread before training. Keep this on when properties have very different scales."},{"location":"reference/config.html#model","title":"<code>model:</code>","text":"setting type default meaning <code>latent_dim</code> integer <code>128</code> Size of the shared latent space where structures and properties meet. <code>node_hidden_dim</code> integer <code>128</code> Width of the per-atom features inside the graph network. <code>edge_dim</code> integer <code>64</code> Width of the per-bond messages inside the graph network. <code>species_embedding_dim</code> integer <code>64</code> Size of the learned element embedding. <code>property_hidden_dim</code> integer <code>128</code> Width of the property encoder's hidden layer. <code>decoder_coordinate_input</code> <code>data</code> / <code>zeros</code> / <code>prototype</code> <code>data</code> What the crystal decoder receives as starting atom positions during training. 'data' (published model) uses the true positions; generation always starts from 'zeros' or the prototype. 'zeros'/'prototype' make training and generation consistent (experimental)."},{"location":"reference/config.html#training","title":"<code>training:</code>","text":"setting type default meaning <code>epochs</code> integer <code>200</code> Number of passes over the training data. <code>batch_size</code> integer <code>16</code> Materials per training step (use 8 if memory is short). <code>learning_rate</code> number <code>0.001</code> Adam learning rate. <code>contrastive_weight</code> number <code>5.0</code> Strength of the alignment between structure and property latents. <code>temperature</code> number <code>0.01</code> Sharpness of the contrastive (InfoNCE) alignment. <code>contrastive_warmup_epochs</code> integer/null <code>None</code> Epochs over which the alignment strength ramps up from 0. Default: 60% of the epochs (the published model used 1200 of 2000). <code>loss_weights</code> section <code>seed</code> integer <code>0</code> Random seed for weight initialisation and data shuffling. <code>device</code> string <code>auto</code> 'auto', 'cpu' or 'cuda'. <code>save_every</code> integer <code>50</code> Write an intermediate checkpoint every N epochs."},{"location":"reference/config.html#trainingloss_weights","title":"<code>training.loss_weights:</code>","text":"setting type default meaning <code>joint_reconstruction</code> number <code>1.0</code> Rebuild the crystal from the joint (structure+property) latent. <code>joint_property</code> number <code>1.0</code> Predict the properties from the joint latent. <code>property_reconstruction</code> number <code>0.5</code> Rebuild the crystal from the property latent alone - this is what makes inverse design possible. <code>property_property</code> number <code>0.5</code> Predict the properties back from the property latent."},{"location":"reference/config.html#generation","title":"<code>generation:</code>","text":"setting type default meaning <code>family</code> string <code>perovskite_abx3</code> Built-in family name or path to your own family .yaml file. <code>variant</code> string/null <code>None</code> Sub-family, e.g. 'halide' or 'oxide' for perovskites. <code>objectives</code> array required Which properties to steer and how. <code>targets</code> array required One entry per design target, e.g. {dir_gap: 1.5, heat_all: -0.1}. Each runs its own search. <code>per_target</code> integer <code>4</code> Candidates to save for each target. <code>population</code> integer <code>48</code> Latent vectors optimised in parallel in each round. <code>rounds</code> integer <code>20</code> Maximum search rounds per target. <code>steps</code> integer <code>800</code> Gradient steps per round. <code>learning_rate</code> number <code>0.0012</code> Step size of the latent search. <code>anchor_weight</code> number <code>12.0</code> Keeps the search close to the latent that the target properties map to. <code>temperature_start</code> number <code>1.6</code> Softness of element choices at the start of a round. <code>temperature_end</code> number <code>0.9</code> Softness of element choices at the end of a round. <code>diversity_weight</code> number <code>0.9</code> Pushes the parallel searches apart from each other. <code>diversity_tau</code> number <code>0.25</code> <code>history_weight</code> number <code>1.0</code> Pushes new searches away from latents that already produced a saved candidate. <code>history_tau</code> number <code>0.25</code> <code>restart_patience</code> integer <code>250</code> Steps without improvement before a search is restarted with noise. <code>restart_noise</code> number <code>0.35</code> <code>grad_clip</code> number <code>1.0</code> <code>z_clip</code> number <code>5.0</code> <code>init_sigma</code> number <code>0.35</code> Noise added to the starting latents. <code>init_anchor_mix</code> number <code>0.75</code> Share of the property-encoder latent in the starting point. <code>init_rff_mix</code> number <code>0.6</code> Share of the target-dependent random-feature direction in the starting point. <code>init_noise_mix</code> number <code>0.25</code> Share of random noise in the starting point. <code>rff_frequencies</code> integer <code>48</code> <code>decode_temperature</code> number <code>1.25</code> Randomness when turning element scores into a choice. <code>decode_topk</code> integer <code>12</code> Only the k best-scoring elements of a group can be chosen. <code>decode_tries</code> integer <code>12</code> Attempts per latent to find a composition that passes all constraints. <code>anti_repeat_alpha</code> number <code>0.6</code> Down-weights elements that were already used in saved candidates. <code>anti_repeat_group</code> string/null <code>None</code> Group the anti-repeat applies to (default: first sampled group). <code>geometry_scale</code> number <code>1.0</code> Global strength of the family's search terms. <code>property_first</code> boolean <code>False</code> Ignore the family's search terms during the search (properties only). <code>dedup_formula</code> boolean <code>True</code> Save each composition at most once. <code>min_cosine_sep</code> number <code>0.985</code> Skip candidates whose latent is this similar (cosine) to an already saved one. <code>unique_decimals</code> integer <code>3</code> <code>exclude_elements</code> array Elements never to use, e.g. [Pb, Cd]. <code>only_elements</code> object Restrict groups to these elements, e.g. {B: [Ti, Zr, Hf]}. <code>overrides</code> object Change parameters of the family's search terms or constraints, e.g. {tolerance_factor: {max: 1.0}}. <code>extra_constraints</code> array Additional rules appended to the family's constraints, e.g. [{name: property_window, property: dir_gap, min: 1.0, max: 3.0}]. <code>disabled_rules</code> array Rules to switch off for this run, by name (or id when two rules share a name), e.g. [charge_neutrality]. The generation report lists them. <code>seed</code> integer <code>937</code> Random seed of the search. <code>amp</code> boolean <code>True</code> Use mixed precision on GPUs. <code>output_prefix</code> string/null <code>None</code> File-name prefix of saved CIFs (default: family variant)."},{"location":"reference/config.html#generationobjectives","title":"<code>generation.objectives[]</code>","text":"setting type default meaning <code>property</code> string required Property (column name) this objective refers to. <code>loss</code> <code>l2</code> / <code>l1</code> / <code>at_most</code> / <code>at_least</code> <code>l2</code> How a prediction is compared with the target: l2 = squared distance, l1 = absolute distance, at_most = only values above the target are penalised, at_least = only values below. <code>weight</code> number <code>10000.0</code> Importance of this objective during the latent search. <code>select_weight</code> number <code>1.0</code> Importance of this objective when ranking finished candidates. <code>select_loss</code> string/null <code>None</code> Comparison used for ranking (defaults to 'loss')."},{"location":"reference/families.html","title":"Family files","text":"<p>A family file is YAML with these keys. <code>meidnet families NAME</code> prints a summary and the constraints with their explanations; errors name the offending key or element.</p> key required meaning <code>name</code>, <code>title</code>, <code>description</code> name identification; <code>description</code> appears in the Studio <code>prototype.lattice</code> yes <code>cubic</code>, <code>fcc_primitive</code>, <code>bcc_primitive</code>, or explicit <code>a, b, c, alpha, beta, gamma</code> <code>prototype.reference_a</code> typical cubic cell edge (\u00c5), used by the <code>cubic_cell</code> search term <code>prototype.sites[]</code> yes <code>{group, frac}</code> in decoder slot order; fractional coordinates <code>groups.&lt;G&gt;.description</code> sentence shown in reports <code>groups.&lt;G&gt;.elements</code> yes every element the group may hold (all variants) <code>groups.&lt;G&gt;.sample</code> subset offered to the final element choice (default: <code>elements</code>) <code>groups.&lt;G&gt;.oxidation_states</code> <code>{element: [charges]}</code> used for charge balance and neutrality bias <code>sampling_order</code> order in which groups are chosen; the last is restricted to charge-balancing elements <code>lattice.rule</code> <code>bond_sum</code> (<code>a = factor \u00d7 \u03a3 r_ionic</code> of <code>groups</code>, clipped to <code>min..max</code>) or <code>fixed</code> <code>variants.&lt;V&gt;.groups</code> per-variant <code>elements</code>, <code>sample</code>, <code>oxidation_states</code> of any group <code>variants.&lt;V&gt;.params</code> <code>{rule_or_term_name: {param: value}}</code> applied to search terms and constraints <code>require_variant</code> default true: a variant must be chosen when variants exist <code>search_terms[]</code> <code>{name, \u2026params}</code> soft penalties during the latent search <code>logit_transforms[]</code> <code>{name, \u2026params}</code> edits of the decoder's element scores (e.g. <code>neutrality_bias</code>) <code>constraints[]</code> <code>{name, \u2026params}</code> hard rules; <code>min_distance</code> runs before symmetry refinement, the rest after <code>refine_symmetry</code>, <code>symprec</code> run spglib refinement on candidates (default true, 0.05) <p>User-side changes without editing the file: <code>generation.variant</code>, <code>exclude_elements</code>, <code>only_elements</code>, <code>overrides</code> (<code>{name: {param: value}}</code> for any search term or constraint).</p> <p>The built-in files are the complete reference: <code>perovskite_abx3.yaml</code>, <code>double_perovskite_a2bbx6.yaml</code>.</p>"},{"location":"reference/rules.html","title":"Rules and search terms","text":"<p>Rules (hard constraints) decide whether a candidate is saved. Search terms (soft penalties) steer the latent search towards decodable, sensible outputs. Both are referenced by name in family files.</p>"},{"location":"reference/rules.html#built-in-rules","title":"Built-in rules","text":""},{"location":"reference/rules.html#bond_window-sensible-fromto-bonds","title":"<code>bond_window</code> \u2014 Sensible {from}\u2013{to} bonds","text":"<p>Every {from} atom needs a {to} neighbour at {low}\u2013{high} \u00d7 the sum of their ionic radii: bonds may be neither squashed nor stretched.</p>"},{"location":"reference/rules.html#charge_neutrality-charge-balance","title":"<code>charge_neutrality</code> \u2014 Charge balance","text":"<p>The formal charges must add up to zero for at least one combination of the listed oxidation states (e.g. Cs\u207a Pb\u00b2\u207a I\u207b\u00d73 = 0).</p>"},{"location":"reference/rules.html#min_distance-no-overlapping-atoms","title":"<code>min_distance</code> \u2014 No overlapping atoms","text":"<p>Two atoms closer than {min} \u00c5 would overlap, which is physically impossible.</p>"},{"location":"reference/rules.html#octahedral_factor-octahedral-factor","title":"<code>octahedral_factor</code> \u2014 Octahedral factor","text":"<p>\u03bc = r_B / r_X says whether the B cation is large enough to hold six X anions around it ({min} \u2264 \u03bc \u2264 {max}).</p>"},{"location":"reference/rules.html#property_window-predicted-property-window","title":"<code>property_window</code> \u2014 Predicted {property} window","text":"<p>The model's predicted {property} must lie between {min} and {max}.</p>"},{"location":"reference/rules.html#tolerance_factor-goldschmidt-tolerance-factor","title":"<code>tolerance_factor</code> \u2014 Goldschmidt tolerance factor","text":"<p>t = (r_A + r_X) / (\u221a2 (r_B + r_X)) measures how well the A cation fits the cage of B\u2013X octahedra. Cubic perovskites form roughly for {min} \u2264 t \u2264 {max}.</p>"},{"location":"reference/rules.html#search-terms","title":"Search terms","text":"<ul> <li><code>cubic_cell</code> \u2014 Pulls the decoded cell towards a cube of the family's reference volume.</li> <li><code>entropy_bonus</code> \u2014 Rewards uncertainty early in a round so the search explores (fades to zero).</li> <li><code>group_consistency</code> \u2014 Sites of the same group should agree on their element (ramps up during a round).</li> <li><code>prototype_alignment</code> \u2014 Pulls decoded positions onto the prototype's site positions.</li> <li><code>short_distance_penalty</code> \u2014 Large penalty if any two decoded atoms come closer than <code>min</code>.</li> <li><code>site_repulsion</code> \u2014 Short-range exponential repulsion between decoded atoms.</li> <li><code>site_separation</code> \u2014 Keeps decoded atoms from collapsing onto each other (1/distance).</li> <li><code>soft_charge</code> \u2014 Expected total charge of the most likely composition should be zero (ramps up during a round).</li> <li><code>tolerance_penalty</code> \u2014 Soft version of the Goldschmidt tolerance-factor window, using the most likely elements.</li> </ul>"},{"location":"reference/rules.html#logit-transforms","title":"Logit transforms","text":"<ul> <li><code>neutrality_bias</code> \u2014 Raises the scores of anions whose charge can neutralise the most likely cations.</li> </ul>"},{"location":"reference/rules.html#your-own-rule","title":"Your own rule","text":"<pre><code>from meidnet.constraints import CONSTRAINTS\n\n@CONSTRAINTS.register(\"no_lead\", \"Rejects any composition containing Pb.\")\ndef no_lead(cand):\n    return cand.result(\"no_lead\", \"Pb\" not in cand.elements.values())\n</code></pre> <p>Save it as <code>my_rules.py</code>, list it under <code>plugins:</code> in <code>meidnet.yaml</code>, and add <code>- {name: no_lead}</code> to the family's <code>constraints</code>. See Add a rule.</p>"},{"location":"research/paper.html","title":"Paper","text":"<p>MEIDNet: Multimodal generative AI framework for inverse materials design. Anand Babu, Rog\u00e9rio Almeida Gouv\u00eaa, Pierre Vandergheynst, Gian-Marco Rignanese. npj Computational Materials (2026). doi:10.1038/s41524-026-02153-3 \u00b7 arXiv:2601.22009</p> <pre><code>@article{meidnet2026,\n  title   = {MEIDNet: Multimodal generative AI framework for inverse materials design},\n  author  = {Anand Babu and Rog{\\'e}rio Almeida Gouv{\\^e}a and Pierre Vandergheynst and Gian-Marco Rignanese},\n  journal = {npj Computational Materials},\n  year    = {2026},\n  doi     = {10.1038/s41524-026-02153-3}\n}\n</code></pre>"},{"location":"research/paper.html#code-and-data-of-the-paper","title":"Code and data of the paper","text":"<ul> <li>Code as published: tag <code>v1.0.0-paper</code>   (frozen; MEIDNet 2 reproduces it bit for bit \u2014 see Perov-5).</li> <li>Checkpoints: <code>checkpoints/</code> in the repository (2.8 MB each).</li> <li>Data: Perov-5 in the CDVAE split (<code>meidnet download-data</code>); Castelli et al., Energy Environ. Sci. 5, 5814 (2012);   Xie et al., ICLR (2022).</li> <li>Generated structures: <code>examples/perov5/paper_results/</code>.</li> </ul>"},{"location":"research/paper.html#related","title":"Related","text":"<ul> <li>Satorras, Hoogeboom, Welling, E(n) equivariant graph neural networks, ICML 2021 (the encoder).</li> <li>Radford et al., CLIP, ICML 2021 (the alignment objective).</li> <li>Batatia et al., MACE-MP-0 (stability screening).</li> </ul>"},{"location":"start/colab.html","title":"Run in Google Colab","text":"<p>Each notebook is a complete, runnable workflow. Training runs in your Colab session (free GPU is enough): your data never leaves your Google account.</p> Notebook What it does Time 01 \u2014 Quickstart published model, pick a target, read the report, explore the whole design space in 3D 5 min, CPU 02 \u2014 Your own data upload table + CIFs \u2192 check \u2192 train \u2192 generate \u2192 download everything 15 min + training 03 \u2014 Add a rule write a custom constraint in a few lines and see its effect before and after 5 min 04 \u2014 MEIDNet Studio in Colab the live Studio inside the notebook, then the same blocks driven from Python 10 min"},{"location":"start/colab.html#block-by-block-like-the-studio","title":"Block by block, like the Studio","text":"<p>Every notebook follows the seven blocks of the Studio, in order and in the same colours: Block 1 \u00b7 Data (blue), Block 2 \u00b7 Model (orange), Block 3 \u00b7 Family (green), Block 4 \u00b7 Rules (amber), Block 5 \u00b7 Targets (pink), Block 6 \u00b7 Search (dark green) and Block 7 \u00b7 Candidates (violet). Each block starts with a coloured header saying what it does, has a form for its settings, shows its result as charts (and crystals in 3D where it helps), and ends by printing what flows on to the next block.</p> <p>Run the cells from top to bottom (Runtime \u2192 Run all also works). Change a value in a form and re-run from that block on: the later blocks pick up the change.</p> <p>The notebooks are generated from <code>scripts/make_notebooks.py</code>, so they stay in step with the command line.</p>"},{"location":"start/quickstart.html","title":"5-minute quickstart","text":"In the browser (nothing to install)On your computerIn Colab <p>Open the live Studio. It runs the published perovskite model: move rule limits and targets and watch what passes, run a latent search and read each candidate's checklist, export the <code>meidnet.yaml</code>. (Search budgets are capped on the shared server; the local install below has no limits and takes your own data.)</p> <pre><code>pip install meidnet          # PyTorch CPU wheels are fine; CUDA is optional\nmeidnet demo                 # halide perovskites, band gap 2.0 eV \u2192 runs/demo/\n</code></pre> <p><code>demo</code> downloads the 2.8 MB published checkpoint, searches for 3 candidates (about a minute on a CPU), checks them against the family's rules and opens <code>generation_report.html</code>.</p> <pre><code>meidnet demo --family oxide --band-gap 3.0 --enthalpy -0.2 -n 5\nmeidnet studio               # the live workbench with the same model\n</code></pre> <p>Notebook 01 \u2014 quickstart</p>"},{"location":"start/quickstart.html#what-you-just-saw","title":"What you just saw","text":"<ol> <li>Targets \u2192 latent. The property encoder maps (band gap, enthalpy) to a point in the shared latent space.</li> <li>Search. A population of latents near that point is optimised so that the decoder's output matches the    targets and the family's soft terms (cubic cell, sensible charges, \u2026).</li> <li>Decode + rules. Each latent is decoded into one element per site; the composition is placed on the    family's prototype and must pass every hard rule (charge balance, tolerance factor, \u2026).</li> <li>Report. The best new, unique candidates are saved as CIFs with a card that shows each rule's value and    a gauge of predicted vs target property.</li> </ol>"},{"location":"start/quickstart.html#next","title":"Next","text":"<ul> <li>What data do I need? then Bring your own dataset.</li> <li>MEIDNet Studio to see the effect of every change live.</li> </ul>"},{"location":"start/what-data.html","title":"What data do I need?","text":""},{"location":"start/what-data.html#the-checklist","title":"The checklist","text":"You need Format Notes Structures CIF text in a <code>cif</code> column of the table, or one <code>&lt;id&gt;.cif</code> file per material in a folder up to <code>max_sites</code> atoms per cell (default 20); any elements Properties one or more numeric columns any units; each property is standardised unless you say otherwise An id a column whose values match the CIF file names (if you use files) <code>12</code>, <code>12.0</code> and <code>\"12\"</code> are treated as the same id A family the prototype your structures share (e.g. cubic ABX\u2083) needed for generation; training without one is possible (<code>align_to_prototype: false</code>) <p>A table like this is enough:</p> <pre><code>material_id,band_gap,dielectric,cif\nmat_001,1.42,18.3,\"# generated using pymatgen\ndata_CsPbBr3\n_cell_length_a 5.87 ...\"\nmat_002,2.16,11.7,\"...\"\n</code></pre> <p>CSV, Excel (<code>.xlsx</code>) and JSON tables are read; Parquet too once <code>pyarrow</code> is installed (<code>pip install \"meidnet[parquet]\"</code>).</p>"},{"location":"start/what-data.html#how-much-data","title":"How much data?","text":"<p>The published model used 11k structures. A few thousand is typical; a few hundred lets you try the pipeline, but expect rough predictions (the training report tells you how rough).</p>"},{"location":"start/what-data.html#what-happens-to-each-row","title":"What happens to each row","text":"<p><code>meidnet check</code> reads every row exactly as training would and writes <code>check_report.html</code> with one line per reason a row was skipped (unreadable CIF, too many atoms, missing value, does not match the prototype) and example ids. Nothing is silently dropped.</p>"},{"location":"start/what-data.html#properties-and-ranges","title":"Properties and ranges","text":"<p>MEIDNet can only design for property values it has seen. The check report draws each property's distribution; the generation report flags targets outside the training range as extrapolation.</p>"},{"location":"start/what-data.html#which-family","title":"Which family?","text":"<ul> <li>Cubic ABX\u2083 perovskite \u2192 built-in <code>perovskite_abx3</code> (variants: oxide, halide, chalcogenide, nitride).</li> <li>Rock-salt-ordered double perovskite A\u2082BB\u2032X\u2086 \u2192 built-in <code>double_perovskite_a2bbx6</code>.</li> <li>Anything else with a fixed arrangement of sites \u2192 write a family file   (positions, which elements may occupy which site, their charges, a cell-size rule, rules).</li> <li>Mixed prototypes \u2192 you can still train (<code>align_to_prototype: false</code>) and use the model to predict/screen, but   generation needs one prototype.</li> </ul>"},{"location":"understand/how-it-works.html","title":"How MEIDNet works","text":""},{"location":"understand/how-it-works.html#one-latent-space-for-two-modalities","title":"One latent space for two modalities","text":"<pre><code>crystal (CIF) \u2500\u2500\u25ba SE(3)-equivariant encoder \u2500\u2500\u25ba z_c \u2500\u2510                     \u250c\u2500\u25ba crystal decoder   \u2500\u2500\u25ba lattice, species, positions\n                                                      \u251c\u2500\u2500 shared space \u2500\u2500\u2500\u2500\u2500\u2524\nproperties \u2500\u2500\u2500\u2500\u25ba property encoder (MLP) \u2500\u2500\u2500\u2500\u2500\u25ba z_p \u2500\u2518  pulled together     \u2514\u2500\u25ba property decoder  \u2500\u2500\u25ba properties\n                                                        (InfoNCE)\n</code></pre> <ul> <li>The crystal encoder embeds each atom's element, passes messages along neighbour bonds using only squared   distances (so rotating or translating the cell changes nothing), pools the atoms and adds the lattice.</li> <li>The property encoder maps the vector of (normalised) properties to the same space.</li> <li>Both are projected and normalised to unit length; a contrastive loss makes each structure's latent closest   to its own property latent and far from the others in the batch. Its weight ramps up during training so the   decoders stabilise first.</li> <li>The decoders are trained from the joint latent and from the property latent alone \u2014 that second   term is what makes inverse design possible: a point chosen from properties must decode into a structure.</li> </ul> <p>The training objective is</p> \\[ L = w_1 L_\\text{recon}(z_\\text{joint}) + w_2 L_\\text{prop}(z_\\text{joint}) + w_3 L_\\text{recon}(z_p) + w_4 L_\\text{prop}(z_p) + \\lambda(t)\\, L_\\text{InfoNCE}(z_c, z_p). \\]"},{"location":"understand/how-it-works.html#inverse-design-search-in-that-space","title":"Inverse design = search in that space","text":"<p>For each target:</p> <ol> <li>Start a population of latents at the property encoder's output for the target, plus a target-dependent    random direction (so different targets start apart) and noise.</li> <li>Optimise them by gradient descent on: distance of the decoded properties to the target (each objective    with its <code>loss</code> and <code>weight</code>); a cosine anchor to the target latent; the family's search terms (atoms    apart, cubic cell, positions near the prototype, expected charge zero, same element on equivalent sites,    loose tolerance factor); and repulsion between population members and from earlier successes (diversity).</li> <li>Decode: for each latent, choose one element per site group from the decoder's scores with temperature    and top-k sampling; the last group is drawn only from elements whose charge balances the others. Place the    elements on the prototype; the cell edge comes from ionic radii (<code>a = 2(r_B + r_X)</code> for perovskites).</li> <li>Rules: symmetry-refine, then apply every hard constraint. Only compositions that pass all are kept.</li> <li>Rank by distance to the target (<code>select_weight</code>, <code>select_loss</code>), skip duplicates and near-duplicate    latents, save as CIF.</li> </ol>"},{"location":"understand/how-it-works.html#what-family-means","title":"What \"family\" means","text":"<p>A family is data, not code: a prototype (site positions), site groups (allowed elements and oxidation states), a cell-size rule, soft search terms, and hard rules. The published perovskite tables are <code>families/perovskite_abx3.yaml</code>; the same machinery runs <code>double_perovskite_a2bbx6.yaml</code> and any file you write.</p>"},{"location":"understand/how-it-works.html#the-design-space","title":"The design space","text":"<p>For a prototype family, the set of compositions the decoder can output is finite. <code>meidnet space</code> and the Studio enumerate it and score each composition with the structure encoder. The search's candidates are a subset of this space; comparing the two views (encoder prediction vs search's decoded prediction) is a useful sanity check that the reports encourage.</p>"},{"location":"understand/how-it-works.html#reproducibility","title":"Reproducibility","text":"<p>MEIDNet 2 reproduces MEIDNet 1 exactly: the same featurisation, forward pass, training step and generation output (<code>tests/test_v1_regression.py</code> compares against the frozen v1 code in <code>tests/legacy_v1/</code>). One fix: v1's results depended on Python's hash seed through set iteration order; v2 fixes element order, so the same seed gives the same CIFs in any process.</p>"},{"location":"understand/interpretability.html","title":"Why you can trust (or not) a candidate","text":"<p>MEIDNet shows its reasoning rather than a score. Where to look:</p>"},{"location":"understand/interpretability.html#before-trusting-the-model","title":"Before trusting the model","text":"<ul> <li>Training report \u2192 accuracy table. A property rated weak means its predictions are not much better than   guessing the mean; targets on it are unreliable.</li> <li>Training report \u2192 alignment. Low retrieval accuracy means property latents do not land near the right   structures; inverse design from them is guesswork.</li> <li>Check report \u2192 distributions. A target in a region with almost no training data is extrapolation.</li> </ul>"},{"location":"understand/interpretability.html#for-each-candidate","title":"For each candidate","text":"<ul> <li>The checklist \u2014 every rule's measured value and allowed window. These are deterministic chemistry   (charges, radii, distances), not model outputs.</li> <li>The gauges \u2014 where the prediction sits relative to the target and to the training range.</li> <li>Warnings \u2014 predicted value outside the training range (extrapolation) and latent hit the search limit   (the search pushed a latent to the <code>z_clip</code> bound; predictions there are less reliable). The model is trained   on latents of length 1 and every search moves them further out (typically to length 3\u20135), so the report says   this once for the whole run rather than on every card; each candidate's length is the <code>latent_norm</code> column of   <code>candidates.csv</code>.</li> <li>Encoder vs search \u2014 in the Studio, compare the search's decoded prediction with the structure encoder's   prediction for the same composition. Large disagreement is a red flag.</li> </ul>"},{"location":"understand/interpretability.html#for-the-run-as-a-whole","title":"For the run as a whole","text":"<ul> <li>The funnel \u2014 which rule removed most attempts. If it is a chemical rule (charge balance, tolerance   factor), the family's element lists may be too broad for the target; if it is no charge-balancing element,   the oxidation-state tables may be missing an entry.</li> <li>Rejected examples \u2014 what the search wanted to make; often more informative than what it kept.</li> <li>Latent map \u2014 whether the population spread out (diversity) or collapsed.</li> </ul>"},{"location":"understand/interpretability.html#what-meidnet-never-claims","title":"What MEIDNet never claims","text":"<p>A candidate is a composition on an ideal prototype with model-predicted properties. Whether it is stable, synthesisable, or has those properties is for <code>meidnet screen</code>, DFT and the lab to decide.</p>"},{"location":"understand/limits.html","title":"Scope and roadmap","text":""},{"location":"understand/limits.html#what-works-today","title":"What works today","text":"Structures any CIF with up to <code>max_sites</code> atoms (default 20); any elements Properties any number of scalar columns Training your data, your properties; the published settings by default Generation prototype families: a fixed arrangement of sites whose occupants and cell size are chosen. Built in: cubic ABX\u2083, A\u2082BB\u2032X\u2086; write your own in YAML Rules charge balance, tolerance and octahedral factors, bond windows, minimum distance, predicted-property windows, element filters, and your own Python rules Explanations reports for data, training and generation; the Studio for live what-if"},{"location":"understand/limits.html#what-does-not-work-yet","title":"What does not work (yet)","text":"<ul> <li>New atomic arrangements. The decoder's positions are not used to create geometry; candidates are   compositions placed on the family prototype. Generating free geometry needs a decoder trained to produce it   (the published decoder receives the true positions during training).</li> <li>Spectra, images, text as modalities. The property modality is a vector of scalars. Binned XRD or DOS   could be added as a vector modality with its own encoder; it is designed for but not implemented.</li> <li>Large cells. <code>max_sites</code> can be raised, but memory and time grow with the square of it, and the   published model was trained on 5-atom cells.</li> <li>Fine-tuning the published model on new properties. Its property head has two outputs; train a new model.</li> </ul>"},{"location":"understand/limits.html#roadmap-in-order-of-likely-value","title":"Roadmap (in order of likely value)","text":"<ol> <li>Vector modalities (XRD, DOS) with a shared-latent encoder each.</li> <li>A <code>training.init_from</code> option to continue training from a checkpoint with the same properties.</li> <li>A geometry-generating decoder variant (trained from prototype positions instead of true positions \u2014    <code>model.decoder_coordinate_input: prototype</code> is the first step and is available as an experimental setting).</li> <li>More built-in families (spinel, Heusler, rocksalt/zincblende, MXene M\u2082XT\u2082).</li> </ol> <p>Contributions of family files and rules are the easiest way to help; see the GitHub repository.</p>"},{"location":"use/config.html","title":"The configuration file","text":"<p><code>meidnet.yaml</code> holds everything a run depends on. One file = one reproducible experiment: it is copied into every output folder and stored inside <code>model.pt</code>.</p> <pre><code>name: my_oxides                 # outputs \u2192 runs/my_oxides/\nplugins: []                     # Python files with custom rules (optional)\n\ndata:        # where the structures and properties come from        \u2192 meidnet check\nmodel:       # size of the network (defaults are the paper's)\ntraining:    # epochs, batch size, learning rate, alignment schedule \u2192 meidnet train\ngeneration:  # family, objectives, targets, rules overrides, budget  \u2192 meidnet generate\n</code></pre> <p>Rules of thumb:</p> <ul> <li>Unknown keys are errors. A typo such as <code>epocs:</code> stops the run with a message naming the key.</li> <li>Paths are relative to the YAML file, so a project folder can be moved or shared.</li> <li>Defaults reproduce the paper. Leave a section out and the published settings apply.</li> <li><code>meidnet schema</code> prints the JSON Schema (editors such as VS Code use it for completion and validation).</li> </ul> <p>The complete list of settings with their meaning and defaults is in the configuration reference.</p>"},{"location":"use/config.html#the-four-sections-in-one-picture","title":"The four sections in one picture","text":"<pre><code>data:\n  table: materials.csv            # one row per material\n  properties: [{column: band_gap, unit: eV}, {column: dielectric}]\n  align_to_prototype: true        # atoms re-ordered to the family prototype\n\ntraining:\n  epochs: 200\n  contrastive_weight: 5.0         # how strongly structure and property latents are pulled together\n\ngeneration:\n  family: perovskite_abx3\n  variant: oxide\n  objectives:                     # what to steer, and how to compare prediction with target\n    - {property: band_gap, loss: l2, weight: 10000}\n  targets:                        # one search per entry\n    - {band_gap: 2.0}\n  overrides:                      # change a family rule's parameters without editing the family file\n    tolerance_factor: {min: 0.85, max: 1.0}\n  exclude_elements: [Pb]\n</code></pre>"},{"location":"use/explore-3d.html","title":"Explore in 3D","text":"<p>The Studio can show any set of structures as a map linked to a 3D viewer: each point on the map is one crystal; click a point and its structure appears next to the map, where you can rotate and zoom it. The viewer is chemiscope, an open-source structure\u2013property explorer.</p> <p>Open it from any button marked \ud83e\uddca:</p> Button Where Shows \ud83e\uddca Explore all compositions in 3D Family block the design space: every composition of the family on its prototype \ud83e\uddca Map + 3D of all compositions Targets block the same design space, to compare predictions with your targets \ud83e\uddca All candidates in 3D Candidates block the candidates of your last search \ud83e\uddca See these structures in 3D Data block your data after Check my data, or the Perov-5 training data behind the published model \ud83e\uddca show in the 3D explorer the details window of any composition the design space, with that composition selected <p>You can also open the page directly in the explorer with a deep link: <code>?explore=space</code>, <code>?explore=candidates</code> or <code>?explore=data</code>.</p>"},{"location":"use/explore-3d.html#reading-the-map","title":"Reading the map","text":"<ul> <li>Axes. For the design space, the two predicted properties (for the published model: formation enthalpy   across, band gap up). For your data, your property columns. For candidates, their predicted properties. Any   other property can be put on an axis from the map's settings menu.</li> <li>Colour (design space): yellow = passes every rule, purple = rejected by at least one rule. Symbol:   pass / rejected. For candidates the colour is the search score (lower = closer to the targets).</li> <li>Properties you can plot or colour by:</li> </ul> Name Meaning <code>formula</code> the composition <code>site_A</code>, <code>site_B</code>, \u2026 the element on each site group <code>lattice_a</code> the cubic cell edge estimated from ionic radii (\u00c5) <code>pred_&lt;property&gt;</code> the model's prediction from the structure encoder, with its unit <code>rule_&lt;name&gt;</code> the value each rule measures, e.g. <code>rule_tolerance_factor</code> <code>passes_all</code>, <code>verdict</code> 1 / \"pass\" when every rule passes (with the limits currently set on the page) <code>score</code>, <code>round</code>, <code>target_&lt;property&gt;</code>, <code>still_passes</code> candidates only: search score, round found, the target, and whether it still passes the current rules <p>The design space shows up to 2,000 compositions: those passing every rule first, then a fixed random sample of the rest. Your data shows the structures that passed Check my data; the Perov-5 training data shows its first 1,500 structures.</p>"},{"location":"use/explore-3d.html#clicking-a-point","title":"Clicking a point","text":"<p>The crystal appears on the right: drag to rotate, scroll to zoom. The panel below lists every property of that structure. Bonds and the unit cell are drawn. Design-space structures are the ideal prototype filled with that composition's elements, at the cell size the family's lattice rule estimates.</p>"},{"location":"use/explore-3d.html#when-chemiscope-cannot-load","title":"When chemiscope cannot load","text":"<p>The explorer is downloaded from a public CDN the first time you open it (about 3 MB). Without internet access, or if the CDN is blocked, the Studio switches to its built-in viewer: a scatter plot of the same two axes (coloured points pass every rule) linked to the Studio's own crystal viewer \u2014 click a point, drag the cell to rotate, double-click to spin. Nothing else changes.</p>"},{"location":"use/explore-3d.html#save-the-dataset","title":"Save the dataset","text":"<p>\u2b07 dataset JSON saves what you see as a chemiscope dataset. Open it later at chemiscope.org (the site reads the file inside your browser; nothing is uploaded), share it with a colleague, or put it in a paper's supplementary material.</p> <p>The same datasets can be produced without the Studio:</p> <pre><code>meidnet space meidnet.yaml --chemiscope space.json   # the design space of your configuration\n</code></pre> <p>and from Python:</p> <pre><code>from meidnet.studio.chemiscope import space_dataset, candidates_dataset, records_dataset\n</code></pre> <p>In the Colab notebooks, the <code>show3d(...)</code> helper displays structures inline with <code>chemiscope.show</code>.</p>"},{"location":"use/explore-3d.html#credits","title":"Credits","text":"<p>chemiscope is a third-party library (BSD-3-Clause licence) by G. Fraux, R. K. Cersonsky and M. Ceriotti. If you use the 3D explorer in published work, please cite: G. Fraux, R. K. Cersonsky, M. Ceriotti, Chemiscope: interactive structure-property explorer for materials and molecules, Journal of Open Source Software 5, 2117 (2020).</p>"},{"location":"use/reports.html","title":"Reading the reports","text":"<p>Each command writes a self-contained HTML file in <code>runs/&lt;name&gt;/</code>. All three open with a one-sentence verdict (green / amber / red), show the evidence, and end with what to do next.</p>"},{"location":"use/reports.html#check_reporthtml","title":"<code>check_report.html</code>","text":"<ul> <li>Verdict: how many rows are usable; the most common problem if many were skipped.</li> <li>What MEIDNet read: table, columns, structure source, family and how many structures were re-ordered to   the prototype.</li> <li>Rows that were skipped: one row per reason with counts and example ids, and a \"how do I fix these?\" box.</li> <li>Your properties: histogram, min/median/max, and warnings (e.g. \"72 % of band-gap values are 0 \u2014 the model   sees few examples of other values\").</li> <li>Your structures: atoms per cell, most common elements.</li> </ul>"},{"location":"use/reports.html#training_reporthtml","title":"<code>training_report.html</code>","text":"<ul> <li>Verdict: are the properties predicted well from structure alone?</li> <li>How accurate: per property, typical error (MAE), natural spread (std), R\u00b2, and a word \u2014 good if the error   is under a quarter of the spread, fair under a half, weak otherwise \u2014 with predicted-vs-true plots.</li> <li>Did the two modalities align: retrieval accuracy (a structure's own property vector is its nearest match   x % of the time) and the cosine curve. Inverse design depends on this.</li> <li>Training curves and notes (e.g. the alignment ramp longer than training).</li> </ul>"},{"location":"use/reports.html#generation_reporthtml","title":"<code>generation_report.html</code>","text":"<ul> <li>Verdict: how many candidates, and whether any predicted value lies outside the training range.</li> <li>Per target: the funnel (compositions alive after each rule \u2014 the biggest drop is the rule that limits the   family most), the candidate cards, and examples of rejected compositions with the rule and the   measured value.</li> <li>A candidate card shows: elements per site, cell edge, round found, a gauge per objective (training range   as a track, target as a tick, prediction as a dot), the checklist with each rule's value and allowed   window (hover for the explanation), and warnings: prediction outside the training range, or a latent far   from where training latents live (predictions less reliable).</li> <li>Where the search went: latents before and after optimisation projected to two dimensions.</li> </ul>"},{"location":"use/reports.html#candidatescsv-and-generationjson","title":"<code>candidates.csv</code> and <code>generation.json</code>","text":"<p>The CSV has one row per candidate with targets, predictions, elements, cell edge, every rule's value, score and file name. The JSON holds the full funnel counts, rejected examples and settings \u2014 everything the report shows.</p>"},{"location":"use/studio.html","title":"MEIDNet Studio \u2014 the live workbench","text":"<pre><code>meidnet studio                 # the published perovskite model\nmeidnet studio meidnet.yaml    # your project\n</code></pre> <p>A page opens in your browser. The workflow is drawn as a strip of seven coloured blocks; click a block to open it. Whatever you change, the blocks after it update at once, and a line under the strip says, in words, what the change did to each later block \u2014 the way a LabVIEW diagram shows data flowing after you turn a knob.</p> <p>No installation? Use the hosted Studio \u2014 the same page, running on a shared server.</p> <pre><code> Data \u2500\u2500\u25ba Model \u2500\u2500\u25ba Family \u2500\u2500\u25ba Rules \u2500\u2500\u25ba Targets \u2500\u2500\u25ba Search \u2500\u2500\u25ba Candidates\n 11,356    MAE       halide     5 rules   Eg 2.0     idle       0 found\n materials 0.17 eV   924 comps  27 pass   \u0394Hf \u22120.1\n</code></pre>"},{"location":"use/studio.html#reading-the-workflow-strip","title":"Reading the workflow strip","text":"<p>Each block has its own colour, used everywhere on the page (its panel, its charts, its chips in the flow chain):</p> Block Colour What it holds Data blue the table of materials the model learns from Model orange the trained network (published or yours) Family green the crystal prototype and the elements allowed on each site Rules amber the hard checks every candidate must pass Targets pink the property values you want Search dark green the latent search of the paper Candidates violet what the search found <ul> <li>The dot in the corner of a block tells its state: grey = idle, green = fine, amber = just changed or needs a   look (for example a target outside the training range), red = error, pulsing = running.</li> <li>The two lines inside a block are its live status: \"924 comps\", \"27 pass all\", \"MAE 0.17\", \"running\".</li> <li>The label on each arrow says what flows to the next block: \"11,356 materials \u00b7 2 properties\",   \"924 compositions\", \"27 pass all rules\", \"3 found\".</li> <li>Moving dots on the arrows appear after every change: they travel from the block you edited to every block   it reaches, and those blocks flash. This is the quickest way to see what depends on what.</li> </ul> <p>Dashed pink lines are always a target; a violet \u2605 is always something the search found; green \u2713 means passes.</p>"},{"location":"use/studio.html#the-flow-chain-and-the-flow-log","title":"The flow chain and the flow log","text":"<p>Under the strip, the flow chain repeats the last change in words, block by block, for example:</p> <p>Rules: Goldschmidt tolerance factor max 1.05 \u2192 1 \u279c Rules: pass all rules 27 \u2192 23 (\u22124) \u279c Targets: best match CsSnBr\u2083 \u2192 CsGeBr\u2083 \u279c Candidates: 1 of 3 found candidates would now be rejected</p> <p>Click a step to open that block.</p> <p>The Flow log below it keeps the last 60 changes. Click an entry to replay its flow on the strip.</p>"},{"location":"use/studio.html#what-is-live-and-how","title":"What is live, and how","text":"<p>For a prototype family the set of possible compositions is finite (924 for halide perovskites, 10,340 for halide double perovskites). When you open a family the server builds this design space once: every composition on the prototype, the value of each rule's descriptor (tolerance factor, octahedral factor, \u2026) and the model's predicted properties from the structure encoder. From then on, the page can answer instantly:</p> You change You see immediately a rule's limit how many compositions still pass that rule and all rules; the histogram of the descriptor with the window shaded; the funnel an element (click a chip to exclude it) compositions removed, new counts, new best matches a target the compositions predicted closest; the target drawn on the predicted-property histogram and scatter; an extrapolation warning if it lies outside the training range the comparison kind (as close as possible / at most / at least) the ranking anything, after a search which of the found candidates would now be rejected <p>Descriptor values are computed by the same Python functions that <code>meidnet generate</code> uses, so what the page shows is what the search enforces.</p>"},{"location":"use/studio.html#the-two-directions-of-the-model","title":"The two directions of the model","text":"<p>The live tables use structure \u2192 property (encode the prototype structure, decode its properties): fast, and it scores every composition. The Search block runs the paper's property \u2192 structure latent optimisation and shows its progress (target, round, step, loss), the log, and each candidate as it is found, with its checklist. Both views are of the same network; comparing them tells you whether the search found what the encoder considers best, and whether the predictions agree.</p>"},{"location":"use/studio.html#the-blocks","title":"The blocks","text":"<ul> <li>Data \u2014 how many materials were usable, why rows were skipped, property distributions, the elements   present; Bring your own data (below).</li> <li>Model \u2014 properties, units, training ranges, validation errors, loss and alignment curves; Train your own   model (below).</li> <li>Family \u2014 prototype sites with a rotatable 3D preview of the cell, element chips per site group with their   charges (click to exclude), cell-size rule.</li> <li>Rules \u2014 each rule with its explanation, parameters, pass count and descriptor histogram; add a   predicted-property window rule; the funnel.</li> <li>Targets \u2014 one card per objective: target value, comparison kind, weights; the predicted-property scatter   of all compositions (coloured = passes, grey = rejected, \u2605 = found by the search). Click any point or table row   for its checklist and its crystal in 3D.</li> <li>Search \u2014 budget (candidates, latents per round, rounds, steps), a rough time estimate, Run/Stop, progress,   log, link to the full report.</li> <li>Candidates \u2014 one card per candidate: a rotatable 3D cell, elements, gauges of predicted values against the   targets, the rule checklist, warnings, CIF links.</li> </ul> <p>Crystals are drawn by a small built-in viewer: drag to rotate, double-click to spin.</p>"},{"location":"use/studio.html#beginner-view-and-researcher-mode","title":"Beginner view and Researcher mode","text":"<p>The page opens in the beginner view: every block has an Input \u279c Logic \u279c Output strip with live numbers (\"924 compositions \u279c 5 hard rules, applied in order \u279c 27 pass every rule\") and the name of the Python function that does the work, a short In plain words explanation, and a glossary in the side panel (composition, rule, target, latent space, extrapolation).</p> <p>The Researcher mode switch in the header trades these for the technical view: the \u270e Edit as text YAML drawer, and a Behind the scenes panel in every block with the <code>meidnet.yaml</code> and the Python of that block. The choice is remembered by your browser.</p>"},{"location":"use/studio.html#behind-the-scenes","title":"Behind the scenes","text":"<p>At the bottom of every block, Behind the scenes shows the same settings as the part of <code>meidnet.yaml</code> they come from and as the Python that does the same thing \u2014 with the values currently on the page. Copy either into your own project; nothing on the page is hidden from the command line.</p>"},{"location":"use/studio.html#bring-your-own-data","title":"Bring your own data","text":"<p>The Data block has an adapter: it maps your column names to what MEIDNet expects. Nothing in your files is changed.</p> <ol> <li>Upload a table \u2014 CSV, Excel (<code>.xlsx</code>) or JSON, one row per material \u2014 with an id column, numeric property    columns and the structures, either as a column of CIF text or as a zip of <code>&lt;id&gt;.cif</code> files (select both files    together). Parquet works too when <code>pyarrow</code> is installed.</li> <li>Map the columns. The Studio suggests the id column, the CIF column and the numeric property columns; pick    the properties you want, give each a unit (eV, eV/atom, \u2026) and optionally a label, and choose the family (and    variant) your structures belong to. Align to prototype re-orders each structure's atoms onto the family's    sites, which generation needs; raise the prototype tolerance if many rows are skipped as \"does not match the    prototype\".</li> <li>Check my data. Every row is parsed and checked; you see how many are usable, why the others were skipped    (with examples), the distribution of each property and the elements present.</li> <li>Train in the Model block: choose the number of epochs and press \u25b6 Train. The loss and alignment curves    grow while it trains; at the end the validation errors appear and every block switches to your model and    your properties \u2014 the design space is re-scored, the targets are your properties. Each property's typical    error (MAE) is judged in words against the spread (standard deviation) of that property in your data \u2014 good below a quarter of    the spread, fair below half, weak above \u2014 the same rule the training report uses; \"matched correctly\" is    how often a validation structure is matched to its own properties.</li> <li>Switch between your model and the published model at any time in the Model block. Start over    removes your upload and model from the session.</li> </ol> <p>A few epochs give a rough model in a minute or two; 100\u2013200 epochs are typical for real work \u2014 run <code>meidnet train</code> locally (with a GPU if you have one) and open the result with <code>meidnet studio meidnet.yaml</code>.</p> <p>On the hosted Studio</p> <p>Each browser tab has its own private session: other visitors never see your data, model or candidates, and a session's files are deleted after an hour without activity. The shared machine allows up to 1,500 rows and 8 MB per upload, 30 training epochs, 20 atoms per cell, and two jobs (training or search) at a time across all visitors; search budgets are capped too. Run the Studio on your own computer to lift every limit.</p>"},{"location":"use/studio.html#edit-the-configuration-as-text","title":"Edit the configuration as text","text":"<p>\u270e Edit as text in the header opens a drawer with the <code>generation</code> section of <code>meidnet.yaml</code> exactly as the blocks have it: family, variant, excluded elements, rule overrides, extra rules, objectives, targets and search budget. Change anything and press Apply \u2014 the blocks update and the flow chain shows what changed in each.</p> <ul> <li>Validate only checks without applying. Errors name the exact setting, for example   <code>steps: Input should be greater than or equal to 1</code>.</li> <li>On the hosted Studio, values above the shared limits are accepted but capped, and a note says so:   <code>rounds 99 \u2192 6 (limit on this shared server)</code>.</li> <li>Predicted-property windows added in the Rules block appear as <code>extra_constraints</code>, so the file round-trips:   apply it, download it, run it with <code>meidnet generate</code>.</li> </ul> <p>\u2b07 Download meidnet.yaml (also in the Search block) writes the whole configuration \u2014 rules switched off on the page are listed under <code>disabled_rules</code>, excluded elements become <code>exclude_elements</code>, changed limits become <code>overrides</code> \u2014 so the command line reproduces exactly what the page shows. Settings the blocks do not show (for example search terms set in your own <code>meidnet.yaml</code>) are kept as they are.</p>"},{"location":"use/studio.html#explore-in-3d","title":"Explore in 3D","text":"<p>Buttons marked \ud83e\uddca open the 3D explorer: a map with one point per structure, linked to a 3D viewer \u2014 the design space of the family, the candidates of a search, or the training data. See Explore in 3D.</p>"},{"location":"use/studio.html#deep-links","title":"Deep links","text":"<p>Add to the address to open the page at a given place:</p> Address ends with Opens <code>?panel=rules</code> the Rules block (any block: <code>data</code>, <code>model</code>, <code>family</code>, <code>rules</code>, <code>targets</code>, <code>search</code>, <code>candidates</code>) <code>?explore=space</code> the 3D explorer on the design space (<code>candidates</code> or <code>data</code> work too) <p>Both can be combined: <code>?panel=targets&amp;explore=space</code>.</p>"},{"location":"use/studio.html#dark-mode","title":"Dark mode","text":"<p>The \ud83c\udf19 / \u2600\ufe0f button in the header switches between the light and the dark theme; your browser remembers the choice.</p>"},{"location":"use/studio.html#static-copy-for-a-website","title":"Static copy for a website","text":"<pre><code>meidnet studio --export-static docs/studio.html\n</code></pre> <p>embeds the state and design spaces into one HTML file. Everything works without a server except uploading, training and running a new search \u2014 useful for a plain web host. The hosted Studio is the full server instead:</p> <pre><code>meidnet studio --public --host 0.0.0.0 --port 7860 --run-root /tmp/runs --docs-dir site\n</code></pre> <p><code>--public</code> gives every visitor a private session and caps the budgets so the shared machine stays responsive; <code>--docs-dir</code> also serves a built copy of this documentation at <code>/docs/</code>.</p>"},{"location":"use/your-data.html","title":"Bring your own dataset","text":"<p>This page takes a table of structures and properties to designed candidates. Every step writes a report that explains what happened; nothing requires Python.</p>"},{"location":"use/your-data.html#0-install","title":"0. Install","text":"<pre><code>pip install meidnet\n</code></pre>"},{"location":"use/your-data.html#1-describe-your-data-meidnet-init","title":"1. Describe your data: <code>meidnet init</code>","text":"<pre><code>meidnet init --name my_oxides --table materials.csv \\\n             --id-column material_id --cif-column cif \\\n             --properties band_gap dielectric \\\n             --family perovskite_abx3 --variant oxide\n</code></pre> <p>This writes a commented <code>meidnet.yaml</code>. Open it: every line says what it does. If your CIFs are files instead of a column, set <code>cif_column: null</code> and <code>structures_dir: structures/</code>.</p> <pre><code>data:\n  table: materials.csv\n  id_column: material_id\n  cif_column: cif\n  properties:\n    - {column: band_gap,   unit: eV}\n    - {column: dielectric, unit: \"\"}\n  val_fraction: 0.1\n  max_sites: 20\n  align_to_prototype: true\n</code></pre>"},{"location":"use/your-data.html#2-is-it-usable-meidnet-check","title":"2. Is it usable? <code>meidnet check</code>","text":"<pre><code>meidnet check meidnet.yaml\n</code></pre> <pre><code>412 of 430 materials usable.\n  skipped     14  does not match the perovskite_abx3 prototype   e.g. m_031, m_118, m_204\n  skipped      4  missing property value                          e.g. m_077, m_300\nreport: runs/my_oxides/check_report.html\n</code></pre> <p>The report shows the verdict, the distribution of each property, the elements present and, for each skip reason, how to fix it (for example raise <code>data.prototype_tolerance</code> for distorted cells).</p>"},{"location":"use/your-data.html#3-learn-the-latent-space-meidnet-train","title":"3. Learn the latent space: <code>meidnet train</code>","text":"<pre><code>meidnet train meidnet.yaml            # 200 epochs by default\nmeidnet train meidnet.yaml --epochs 5 # a quick smoke test first\n</code></pre> <p><code>training_report.html</code> answers three questions in plain words: how accurate is each property prediction (error vs natural spread \u2192 good / fair / weak), did the two modalities align (retrieval accuracy), and what were the training curves. The model is saved as <code>runs/my_oxides/model.pt</code> with the property names, units and normalisation inside; <code>meidnet info runs/my_oxides/model.pt</code> describes it.</p> <p>Time</p> <p>About 21 s per epoch per 11k structures on a laptop GPU; roughly 10\u00d7 slower on CPU. Use Colab's free GPU (notebook 02) if you have none.</p>"},{"location":"use/your-data.html#4-design-meidnet-generate","title":"4. Design: <code>meidnet generate</code>","text":"<p>Set the targets in <code>meidnet.yaml</code>:</p> <pre><code>generation:\n  family: perovskite_abx3\n  variant: oxide\n  objectives:\n    - {property: band_gap,   loss: l2,       weight: 10000}\n    - {property: dielectric, loss: at_least, weight: 5000}\n  targets:\n    - {band_gap: 2.0, dielectric: 20}\n    - {band_gap: 3.0, dielectric: 20}\n  per_target: 4\n  exclude_elements: [Pb, Cd]\n</code></pre> <pre><code>meidnet generate meidnet.yaml\n</code></pre> <p><code>generation_report.html</code> shows, per target, the funnel (how many attempts each rule removed), one card per candidate (elements, cell, each rule's measured value and window, predicted vs target gauges, warnings about extrapolation) and examples of rejected compositions. CIFs and <code>candidates.csv</code> are in <code>runs/my_oxides/generation/</code>.</p>"},{"location":"use/your-data.html#5-explore-live-meidnet-studio","title":"5. Explore live: <code>meidnet studio</code>","text":"<pre><code>meidnet studio meidnet.yaml\n</code></pre> <p>Change a rule's limit, exclude an element or move a target and see immediately how many compositions pass and which are predicted closest; run searches and export the YAML. Studio guide.</p>"},{"location":"use/your-data.html#6-confirm","title":"6. Confirm","text":"<p>Predicted properties are estimates. <code>meidnet screen runs/my_oxides/generation</code> (MACE universal potential, optional extra) gives a first stability filter; DFT or experiment has the last word.</p>"}]}
\ No newline at end of file
+{"config":{"lang":["en"],"separator":"[\\s\\-]+","pipeline":["stopWordFilter"],"fields":{"title":{"boost":1000.0},"text":{"boost":1.0},"tags":{"boost":1000000.0}}},"docs":[{"location":"index.html","title":"Home","text":""},{"location":"index.html#meidnet-prism","title":"MEIDNet Prism","text":"<p>\u2190 MEIDNet Prism home \u00b7 the documentation of the MEIDNet framework and its platform.</p> <p>Learn, build and benchmark multimodal AI for materials discovery. MEIDNet is the reference implementation: it designs crystalline materials from target properties, using your own data and rules, configured in one YAML file or in the Studio.</p> <pre><code>   Learn \u2500\u2500\u25ba Architectures \u2500\u2500\u25ba Build \u2500\u2500\u25ba Datasets \u2500\u2500\u25ba Benchmarks \u2500\u2500\u25ba Community\n   what is   which design      MEIDNet    data that    results with   share and\n   multimodal fits your data   Studio     works        evidence       reproduce\n</code></pre> <p>MEIDNet learns one latent space shared by crystal structures and their properties, then searches it for new materials that hit property targets while obeying the chemical and structural rules of a material family. The published cubic-perovskite model (band gap + formation enthalpy) is one application of it; the framework accepts any table of structures and scalar properties and any prototype family you describe.</p>      Your browser cannot play this video: download it.    MEIDNet in one minute: data, model, family, rules, targets, search and candidates. The same tour opens   on a first visit to the live Studio (the \u201c\u25b6 How it works\u201d button replays it). <ul> <li> <p> Learn multimodality</p> <p>What a modality is, the five challenges of multimodal learning, contrastive learning with a playground, and an interactive map of materials modalities.</p> <p> Learn</p> </li> <li> <p> Choose an architecture</p> <p>Early fusion, late fusion, shared latent spaces, cross-attention and contrastive learning, and an advisor that recommends one for your data and goal.</p> <p> Architecture Atlas</p> </li> <li> <p> Try MEIDNet</p> <p>The live Studio runs in the browser without installation; <code>meidnet demo</code> runs on a laptop CPU.</p> <p> 5-minute quickstart</p> </li> <li> <p> Use MEIDNet on my data</p> <p>A table with an id, property columns and CIFs. Four commands lead from a data check to candidates with reports.</p> <p> Bring your own dataset</p> </li> <li> <p> Change what it does</p> <p>Targets, elements, rules and family are set in one YAML file, or with the controls of MEIDNet Studio, which exports the same file.</p> <p> Recipes</p> </li> <li> <p> Understand it</p> <p>How the two modalities are aligned, how the search works, what the reports mean, and where the limits are.</p> <p> How MEIDNet works</p> </li> </ul>"},{"location":"index.html#the-workflow","title":"The workflow","text":"<pre><code>   Data \u2500\u2500\u25ba Model \u2500\u2500\u25ba Family \u2500\u2500\u25ba Rules \u2500\u2500\u25ba Targets \u2500\u2500\u25ba Search \u2500\u2500\u25ba Candidates\n   table    shared    prototype  hard      property   latent     CIFs + report\n   + CIFs   latent    + sites    checks    goals      optimisation\n</code></pre> <p>Each block is a setting in <code>meidnet.yaml</code>, a node in MEIDNet Studio, where a change in one block updates the blocks after it, and a section in the HTML report that each step writes.</p>"},{"location":"index.html#scope","title":"Scope","text":"<ul> <li>Generation works for prototype families: a fixed arrangement of sites whose occupants and cell size are   chosen (ABX\u2083 perovskites, A\u2082BB\u2032X\u2086 double perovskites, and anything you describe the same way). MEIDNet does   not invent new atomic arrangements. Details and roadmap.</li> <li>Predicted properties are the model's estimates. The reports say when a prediction is an extrapolation.   Confirm candidates with DFT or experiment; <code>meidnet screen</code> is a first filter.</li> </ul>"},{"location":"index.html#paper","title":"Paper","text":"<p>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, MEIDNet: Multimodal generative AI framework for inverse materials design, npj Computational Materials (2026). Citation and benchmarks \u00b7 code as published (v1.0).</p> <p>Further reading: A. Babu, N. M. A. Krishnan, Multimodal and cross-modal learning techniques, APL Machine Learning 4, 030901 (2026), doi:10.1063/5.0346744 \u00b7 A. Babu, R. Almeida Gouv\u00eaa, G.-M. Rignanese, Toward automated discovery with generative models multimodal learning and closed loop workflows in inverse materials design, Cell Reports Physical Science 7, 103561 (2026), doi:10.1016/j.xcrp.2026.103561.</p>"},{"location":"benchmarks/index.html","title":"MEIDNet Benchmarks","text":"<p>Reproducible results for multimodal materials representation and inverse design, one leaderboard per dataset. Different datasets are different tasks, so there is no ranking across them: compare rows only within a page. The idea follows Matbench Discovery: fixed data, inspectable submissions, plots and an evidence level per row.</p> 3datasets5results2reproduced here1DFT or experiment"},{"location":"benchmarks/index.html#datasets","title":"Datasets","text":"dataset role results reproduced highest evidence Carbon-24 published: structure-representation generalization (not a multimodal benchmark) 0 0 \u2013 MP-20 published: structure-representation generalization (not a multimodal benchmark) 0 0 \u2013 Perov-5 the published multimodal benchmark: alignment, property reconstruction and inverse design 5 2 DFT validated"},{"location":"benchmarks/index.html#the-evidence-ladder","title":"The evidence ladder","text":"level means \u25cb\u25cb\u25cb\u25cb\u25cb generated the model produced it \u25cf\u25cb\u25cb\u25cb\u25cb ML filtered it passed the family's rules and the model's own checks \u25cf\u25cf\u25cb\u25cb\u25cb MLIP validated relaxed and found stable with a machine-learned potential (<code>meidnet screen</code>, MACE) \u25cf\u25cf\u25cf\u25cb\u25cb DFT validated confirmed by density-functional calculations \u25cf\u25cf\u25cf\u25cf\u25cb experimentally validated synthesised and measured results per evidence level (all datasets)generated3MLIP validated1DFT validated1"},{"location":"benchmarks/index.html#status-of-a-row","title":"Status of a row","text":"<p>Published = quoted from the paper \u00b7 Community submitted = a pull request, validated by CI \u00b7 MEIDNet verified = reproduced here from the submitted configuration (<code>python scripts/benchmarks.py reproduce &lt;id&gt;</code>). How to contribute \u00b7 Databases by application</p>"},{"location":"benchmarks/index.html#latest-results","title":"Latest results","text":"date result dataset category evidence status 2026-10-03 MLIP screening of the shipped paper candidates (MACE-MP-0) Perov-5 Validation level \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated Community submitted 2026-10-03 Shipped checkpoint re-evaluated on the training split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-10-03 Shipped checkpoint re-evaluated on the held-out validation split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-05-29 MEIDNet (paper): inverse design of perovskites from property targets Perov-5 Conditional design \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated Published (from the paper) 2026-05-29 MEIDNet (paper): alignment of structure and property latents on Perov-5 Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated Published (from the paper)"},{"location":"benchmarks/carbon24.html","title":"Carbon-24","text":"<p>About 10,000 carbon structures with up to 24 atoms per cell (CDVAE split). In the MEIDNet paper it is the second structure-representation generalization test; no property modality and no checkpoint here.</p> <p>10,153 rows \u00b7 split: CDVAE: 6,091 train / 2,031 val / 2,031 test \u00b7 role in the paper: published: structure-representation generalization (not a multimodal benchmark) \u00b7 dataset source</p> 0results0reproduced here0categories\u2013highest evidence <p>Read a row</p> <p>Rows are compared only within this dataset and category; click a column header to sort. The ladder \u25cf\u25cb\u25cb\u25cb\u25cb \u2026 \u25cf\u25cf\u25cf\u25cf\u25cf is the evidence level (generated \u2192 ML filtered \u2192 MLIP validated \u2192 DFT validated \u2192 experimentally validated). Published = taken from the paper; MEIDNet verified = reproduced here from the submitted configuration.</p> <p>No results yet. Contribute one.</p>"},{"location":"benchmarks/mp20.html","title":"MP-20","text":"<p>About 45,000 Materials Project structures with at most 20 atoms per cell (CDVAE split). In the MEIDNet paper it tests how the crystal encoder and decoder generalise beyond perovskites; no property modality was aligned on it and no MP-20 checkpoint is shipped in this repository.</p> <p>45,231 rows \u00b7 split: CDVAE: 27,136 train / 9,047 val / 9,046 test \u00b7 role in the paper: published: structure-representation generalization (not a multimodal benchmark) \u00b7 dataset source</p> 0results0reproduced here0categories\u2013highest evidence <p>Read a row</p> <p>Rows are compared only within this dataset and category; click a column header to sort. The ladder \u25cf\u25cb\u25cb\u25cb\u25cb \u2026 \u25cf\u25cf\u25cf\u25cf\u25cf is the evidence level (generated \u2192 ML filtered \u2192 MLIP validated \u2192 DFT validated \u2192 experimentally validated). Published = taken from the paper; MEIDNet verified = reproduced here from the submitted configuration.</p> <p>No results yet. Contribute one.</p>"},{"location":"benchmarks/perov5.html","title":"Perov-5","text":"<p>18,928 cubic ABX3 perovskites with relaxed structures, formation enthalpy (heat_all, eV/atom) and direct band gap (dir_gap, eV). The CDVAE split of the Castelli et al. (2012) perovskite set; the dataset behind the published MEIDNet model and the live Studio.</p> <p>18,928 rows \u00b7 split: CDVAE: 11,356 train / 3,787 val / 3,785 test \u00b7 properties: <code>heat_all</code>, <code>dir_gap</code> \u00b7 role in the paper: the published multimodal benchmark: alignment, property reconstruction and inverse design \u00b7 dataset source</p> 5results2reproduced here3categoriesDFT validatedhighest evidence <p>Read a row</p> <p>Rows are compared only within this dataset and category; click a column header to sort. The ladder \u25cf\u25cb\u25cb\u25cb\u25cb \u2026 \u25cf\u25cf\u25cf\u25cf\u25cf is the evidence level (generated \u2192 ML filtered \u2192 MLIP validated \u2192 DFT validated \u2192 experimentally validated). Published = taken from the paper; MEIDNet verified = reproduced here from the submitted configuration.</p>"},{"location":"benchmarks/perov5.html#representation-quality","title":"Representation quality","text":"result split inputs cosine_matched l2_matched retrieval_top1 retrieval_top5 mae_heat_all r2_heat_all mae_dir_gap r2_dir_gap n_evaluated evidence status MEIDNet (paper): alignment of structure and property la\u2026 validation structure + 2 properties 0.97 0.24 \u2013 \u2013 \u2013 \u2013 \u2013 \u2013 \u2013 \u25cf\u25cb\u25cb\u25cb\u25cb generated Published (from the paper) Shipped checkpoint re-evaluated on the training split (\u2026 train structure + 2 properties 0.771 0.676 0.0101 0.0405 0.397 eV/atom 0.486 2.78 eV -29.8 11,356 materials \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified Shipped checkpoint re-evaluated on the held-out validat\u2026 val structure + 2 properties 0.771 0.676 0.0253 0.0908 0.392 eV/atom 0.497 2.79 eV -35.8 3,787 materials \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified How to read the representation quality table <ul> <li><code>cosine_matched</code> \u2014 mean cosine similarity between the structure and property latents of the same material (1 = aligned)</li> <li><code>l2_matched</code> \u2014 mean L2 distance between those two latents (0 = identical) (lower is better)</li> <li><code>retrieval_top1</code> \u2014 fraction of validation materials whose property latent is nearest to its own structure latent</li> <li><code>retrieval_top5</code> \u2014 the same within the five nearest</li> <li><code>mae_heat_all</code> \u2014 mean absolute error of <code>heat_all</code> predicted from the structure alone (physical units) (lower is better)</li> <li><code>r2_heat_all</code> \u2014 coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect</li> <li><code>mae_dir_gap</code> \u2014 mean absolute error of <code>dir_gap</code> predicted from the structure alone (physical units) (lower is better)</li> <li><code>r2_dir_gap</code> \u2014 coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect</li> <li><code>n_evaluated</code> \u2014 materials in the evaluation split</li> </ul> <p>Representation rows describe how well the shared latent space holds structures and properties; they say nothing about whether a generated material is real.</p> cosine_matched (higher is better)MEIDNet (paper): alignment of structur0.97Shipped checkpoint re-evaluated on the0.771Shipped checkpoint re-evaluated on the0.771retrieval_top1 (higher is better)Shipped checkpoint re-evaluated on the0.0253Shipped checkpoint re-evaluated on the0.0101mae_heat_all (lower is better)Shipped checkpoint re-evaluated on the0.397Shipped checkpoint re-evaluated on the0.392mae_dir_gap (lower is better)Shipped checkpoint re-evaluated on the2.79Shipped checkpoint re-evaluated on the2.78dir_gap: predicted from the structure vs. true (3,787 validation materials)02460246true dir_gap (eV)predicted dir_gap heat_all: predicted from the structure vs. true (3,787 validation materials)024024true heat_all (eV/atom)predicted heat_all cosine similarity of the two latents, per material0.650.70.750.80.850200400600800cosine(structure latent, property latent)count0.636\u20130.642: 10.672\u20130.678: 10.689\u20130.695: 10.695\u20130.701: 10.701\u20130.707: 20.707\u20130.713: 30.713\u20130.719: 30.719\u20130.725: 60.725\u20130.731: 50.731\u20130.737: 160.737\u20130.743: 280.743\u20130.748: 520.748\u20130.754: 1080.754\u20130.76: 2970.76\u20130.766: 7360.766\u20130.772: 7540.772\u20130.778: 5820.778\u20130.784: 7710.784\u20130.79: 2510.79\u20130.796: 640.796\u20130.802: 290.802\u20130.808: 290.808\u20130.813: 180.813\u20130.819: 170.819\u20130.825: 50.825\u20130.831: 30.831\u20130.837: 30.855\u20130.861: 1mean"},{"location":"benchmarks/perov5.html#conditional-design","title":"Conditional design","text":"result split inputs n_generated n_sun sun_rate evidence status MEIDNet (paper): inverse design of perovskites from pro\u2026 \u2013 structure + 2 properties 140 structures 19 structures 0.136 \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated Published (from the paper) How to read the conditional design table <ul> <li><code>n_generated</code> \u2014 candidates the search produced</li> <li><code>n_sun</code> \u2014 stable, unique and novel candidates</li> <li><code>sun_rate</code> \u2014 n_sun / n_generated</li> </ul> <p>Design rows are funnels: every number after <code>n_generated</code> is a subset of the one before. A high SUN rate on few candidates is weaker evidence than a lower rate on many.</p> MEIDNet (paper): inverse design of perovskites from property targets: from generated to validatedgenerated140stable, unique, novel19"},{"location":"benchmarks/perov5.html#validation-level","title":"Validation level","text":"result split inputs n_screened evidence status MLIP screening of the shipped paper candidates (MACE-MP\u2026 \u2013 structure + 2 properties 26 structures \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated Community submitted How to read the validation level table <ul> <li><code>n_screened</code> \u2014 structures relaxed with the MLIP</li> </ul> <p>Validation rows count structures that survived each check. <code>n_stable</code> depends on the stability threshold used (0.10 eV/atom above the hull by default).</p>"},{"location":"benchmarks/perov5.html#insights","title":"Insights","text":"<ul> <li>Best <code>cosine_matched</code>: MEIDNet (paper): alignment of structure and property latents on Perov-5 (0.97), published (from the paper).</li> <li>Best <code>l2_matched</code>: MEIDNet (paper): alignment of structure and property latents on Perov-5 (0.24), published (from the paper).</li> <li>Best <code>retrieval_top1</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.0253), meidnet verified.</li> <li>Best <code>retrieval_top5</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.0908), meidnet verified.</li> <li>Best <code>mae_heat_all</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.392 eV/atom), meidnet verified.</li> <li>Best <code>r2_heat_all</code>: Shipped checkpoint re-evaluated on the held-out validation split (this code) (0.497), meidnet verified.</li> <li>Best <code>mae_dir_gap</code>: Shipped checkpoint re-evaluated on the training split (this code) (2.78 eV), meidnet verified.</li> <li>Best <code>r2_dir_gap</code>: Shipped checkpoint re-evaluated on the training split (this code) (-29.8), meidnet verified.</li> <li>Evidence levels present: generated, MLIP validated, DFT validated. 2 of 5 rows have been reproduced here.</li> <li>2 row(s) are quoted from the paper; a verified re-run of the same checkpoint, where it exists, gives the exact numbers for this code version.</li> <li>A re-run of MEIDNet (paper): alignment of structure and property latents on Perov-5 here did not agree (<code>cosine_matched</code> 0.97 \u2192 0.771; <code>l2_matched</code> 0.24 \u2192 0.676). The row keeps its status; see its page for the full comparison.</li> </ul>"},{"location":"benchmarks/results/paper-inverse-design.html","title":"MEIDNet (paper): inverse design of perovskites from property targets","text":"<p>Published (from the paper) \u00b7 \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 1.0 (paper), perovskite_abx3) Category Conditional design Submitted by Anand Babu (UCLouvain) on 2026-05-29 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth"},{"location":"benchmarks/results/paper-inverse-design.html#results","title":"Results","text":"metric value split meaning note <code>n_generated</code> 140 structures candidates the search produced candidates generated from property targets in the campaign reported in the paper <code>n_sun</code> 19 structures stable, unique and novel candidates stable, unique and novel after screening <code>sun_rate</code> 0.136 n_sun / n_generated 19 of 140 MEIDNet (paper): inverse design of perovskites from property targets: from generated to validatedgenerated140stable, unique, novel19"},{"location":"benchmarks/results/paper-inverse-design.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated.</p> <p>Evidence:</p> <ul> <li>https://doi.org/10.1038/s41524-026-02153-3</li> <li><code>examples/perov5/paper_results/</code></li> </ul> <p>These numbers are quoted from the paper.</p>"},{"location":"benchmarks/results/paper-inverse-design.html#notes","title":"Notes","text":"<p>Numbers quoted from the paper. The 26 CIFs in examples/perov5/paper_results/ are the candidates shipped with the paper; the MLIP screening row re-runs <code>meidnet screen</code> on them.</p>"},{"location":"benchmarks/results/paper-representation.html","title":"MEIDNet (paper): alignment of structure and property latents on Perov-5","text":"<p>Published (from the paper) \u00b7 \u25cf\u25cb\u25cb\u25cb\u25cb generated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 1.0 (paper), perovskite_abx3) Category Representation quality Submitted by Anand Babu (UCLouvain) on 2026-05-29 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Notes on the model 128-dimensional shared latent space; 2000 epochs, contrastive warm-up over the first 1200"},{"location":"benchmarks/results/paper-representation.html#results","title":"Results","text":"metric value split meaning note <code>cosine_matched</code> 0.97 validation mean cosine similarity between the structure and property latents of the same material (1 = aligned) mean cosine similarity between the structure and property latents of the same material, as reported in the paper <code>l2_matched</code> 0.24 validation mean L2 distance between those two latents (0 = identical) mean L2 distance between the two latents of the same material, as reported in the paper"},{"location":"benchmarks/results/paper-representation.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cb\u25cb\u25cb\u25cb generated.</p> <p>Evidence:</p> <ul> <li>https://doi.org/10.1038/s41524-026-02153-3</li> </ul>"},{"location":"benchmarks/results/paper-representation.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>, split val:</p> <pre><code>python scripts/benchmarks.py reproduce paper-representation\n</code></pre> <p>These numbers are quoted from the paper.</p>"},{"location":"benchmarks/results/paper-representation.html#re-run-here-did-not-agree","title":"Re-run here did not agree","text":"<p>On 2026-10-03 this repository's code re-ran the recipe above (AMD64 CPU, MEIDNet 2.0.0) and obtained different numbers. The row keeps its status; the comparison is shown so that readers can judge it. Possible reasons: a different split or protocol in the original measurement, a different definition of the metric, or a checkpoint that is not the one the numbers were measured on.</p> metric this row obtained here agree <code>cosine_matched</code> 0.97 0.771 no <code>l2_matched</code> 0.24 0.676 no <p>Also obtained in that run: <code>mae_heat_all</code> = 0.392, <code>r2_heat_all</code> = 0.497, <code>mae_dir_gap</code> = 2.79, <code>r2_dir_gap</code> = -35.8, <code>retrieval_top1</code> = 0.0253, <code>retrieval_top5</code> = 0.0908, <code>n_evaluated</code> = 3,787</p>"},{"location":"benchmarks/results/paper-representation.html#notes","title":"Notes","text":"<p>Numbers quoted from the paper (npj Computational Materials, 2026). <code>reproduce</code> re-evaluates the shipped checkpoint on the CDVAE validation split and adds MAE, R2 and retrieval metrics; cosine_matched is compared with the quoted value.</p>"},{"location":"benchmarks/results/paper-results-mace-screen.html","title":"MLIP screening of the shipped paper candidates (MACE-MP-0)","text":"<p>Community submitted \u00b7 \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 2.0.0, perovskite_abx3) Category Validation level Submitted by Anand Babu (UCLouvain) on 2026-10-03 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Resources laptop CPU"},{"location":"benchmarks/results/paper-results-mace-screen.html#results","title":"Results","text":"metric value split meaning note <code>n_screened</code> 26 structures structures relaxed with the MLIP the CIFs in examples/perov5/paper_results/"},{"location":"benchmarks/results/paper-results-mace-screen.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated.</p> <p>Evidence:</p> <ul> <li><code>examples/perov5/paper_results/</code></li> </ul>"},{"location":"benchmarks/results/paper-results-mace-screen.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>:</p> <pre><code>meidnet screen examples/perov5/paper_results --train-csv data/perov5/train.csv\n</code></pre>"},{"location":"benchmarks/results/paper-results-mace-screen.html#notes","title":"Notes","text":"<p>n_stable / n_unique / n_novel / n_sun are filled in by <code>python scripts/benchmarks.py reproduce paper-results-mace-screen</code> (MACE-MP-0 relaxation, about ten minutes on a CPU).</p>"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html","title":"Shipped checkpoint re-evaluated on the training split (this code)","text":"<p>MEIDNet verified \u00b7 \u25cf\u25cb\u25cb\u25cb\u25cb generated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 2.0.0, perovskite_abx3) Category Representation quality Submitted by Anand Babu (UCLouvain) on 2026-10-03 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Resources laptop CPU (AMD64)"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#results","title":"Results","text":"metric value split meaning note <code>cosine_matched</code> 0.771 train mean cosine similarity between the structure and property latents of the same material (1 = aligned) mean over the 11,356 training materials <code>l2_matched</code> 0.676 train mean L2 distance between those two latents (0 = identical) <code>retrieval_top1</code> 0.0101 train fraction of validation materials whose property latent is nearest to its own structure latent <code>retrieval_top5</code> 0.0405 train the same within the five nearest <code>mae_heat_all</code> 0.397 eV/atom train mean absolute error of <code>heat_all</code> predicted from the structure alone (physical units) <code>r2_heat_all</code> 0.486 train coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>mae_dir_gap</code> 2.78 eV train mean absolute error of <code>dir_gap</code> predicted from the structure alone (physical units) <code>r2_dir_gap</code> -29.8 train coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>n_evaluated</code> 11,356 materials train materials in the evaluation split"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cb\u25cb\u25cb\u25cb generated.</p> <p>Evidence:</p> <ul> <li><code>checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth</code></li> </ul>"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>, split train:</p> <pre><code>python scripts/benchmarks.py reproduce shipped-checkpoint-train-split\n</code></pre>"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#verified-here","title":"Verified here","text":"<p>Reproduced on 2026-10-03 with MEIDNet 2.0.0 by Anand Babu on AMD64 CPU:</p> metric submitted obtained here <code>mae_heat_all</code> 0.397 0.397 <code>r2_heat_all</code> 0.486 0.486 <code>mae_dir_gap</code> 2.78 2.78 <code>r2_dir_gap</code> -29.8 -29.8 <code>retrieval_top1</code> 0.0101 0.0101 <code>retrieval_top5</code> 0.0405 0.0405 <code>cosine_matched</code> 0.771 0.771 <code>n_evaluated</code> 11,356 11,356 <code>l2_matched</code> 0.676 0.676"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#per-material-analysis","title":"Per-material analysis","text":"<p>Every validation material, predicted from its structure alone; the dashed line is perfect agreement.</p> dir_gap: predicted from the structure vs. true (11,356 validation materials)0246802468true dir_gap (eV)predicted dir_gap heat_all: predicted from the structure vs. true (11,356 validation materials)024024true heat_all (eV/atom)predicted heat_all cosine similarity of the two latents, per material00.502,5005,0007,50010,000cosine(structure latent, property latent)count-0.436\u2013-0.4: 10.211\u20130.247: 20.391\u20130.427: 10.463\u20130.499: 10.499\u20130.535: 10.535\u20130.571: 10.607\u20130.643: 30.643\u20130.679: 20.679\u20130.715: 160.715\u20130.751: 4940.751\u20130.787: 101520.787\u20130.823: 6270.823\u20130.859: 480.859\u20130.895: 50.895\u20130.931: 2mean"},{"location":"benchmarks/results/shipped-checkpoint-train-split.html#notes","title":"Notes","text":"<p>The same evaluation on the data the model was trained on. The numbers equal the held-out ones within noise: the gap between the paper's figures and these is not a train / test effect.</p>"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html","title":"Shipped checkpoint re-evaluated on the held-out validation split (this code)","text":"<p>MEIDNet verified \u00b7 \u25cf\u25cb\u25cb\u25cb\u25cb generated</p> Dataset Perov-5 Modalities <code>structure</code>, <code>property:heat_all</code>, <code>property:dir_gap</code> Model MEIDNet early fusion + curriculum (production model) (MEIDNet 2.0.0, perovskite_abx3) Category Representation quality Submitted by Anand Babu (UCLouvain) on 2026-10-03 Configuration <code>examples/perov5/meidnet.yaml</code> Checkpoint dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth Resources laptop CPU (AMD64) Notes on the model the checkpoint shipped with the repository and used by the live Studio; 128-dimensional shared latent space"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#results","title":"Results","text":"metric value split meaning note <code>cosine_matched</code> 0.771 val mean cosine similarity between the structure and property latents of the same material (1 = aligned) mean over the 3,787 validation materials <code>l2_matched</code> 0.676 val mean L2 distance between those two latents (0 = identical) unit latents: sqrt(2 - 2 cos) <code>retrieval_top1</code> 0.0253 val fraction of validation materials whose property latent is nearest to its own structure latent <code>retrieval_top5</code> 0.0908 val the same within the five nearest <code>mae_heat_all</code> 0.392 eV/atom val mean absolute error of <code>heat_all</code> predicted from the structure alone (physical units) predicted from the structure latent alone <code>r2_heat_all</code> 0.497 val coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>mae_dir_gap</code> 2.79 eV val mean absolute error of <code>dir_gap</code> predicted from the structure alone (physical units) 96% of the rows have a direct gap of 0 eV; the structure-only predictor does not capture that <code>r2_dir_gap</code> -35.8 val coefficient of determination of that prediction; 0 = no better than the mean, 1 = perfect <code>n_evaluated</code> 3,787 materials val materials in the evaluation split"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#validation","title":"Validation","text":"<p>Evidence level: \u25cf\u25cb\u25cb\u25cb\u25cb generated.</p> <p>Evidence:</p> <ul> <li><code>checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth</code></li> <li><code>benchmarks/verified/shipped-checkpoint-val-split.predictions.csv</code></li> </ul>"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#reproduce","title":"Reproduce","text":"<p>Configuration <code>examples/perov5/meidnet.yaml</code>, split val:</p> <pre><code>python scripts/benchmarks.py reproduce shipped-checkpoint-val-split\n</code></pre>"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#verified-here","title":"Verified here","text":"<p>Reproduced on 2026-10-03 with MEIDNet 2.0.0 by Anand Babu on AMD64 CPU:</p> metric submitted obtained here <code>mae_heat_all</code> 0.392 0.392 <code>r2_heat_all</code> 0.497 0.497 <code>mae_dir_gap</code> 2.79 2.79 <code>r2_dir_gap</code> -35.8 -35.8 <code>retrieval_top1</code> 0.0253 0.0253 <code>retrieval_top5</code> 0.0908 0.0908 <code>cosine_matched</code> 0.771 0.771 <code>n_evaluated</code> 3,787 3,787 <code>l2_matched</code> 0.676 0.676"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#per-material-analysis","title":"Per-material analysis","text":"<p>Every validation material, predicted from its structure alone; the dashed line is perfect agreement.</p> dir_gap: predicted from the structure vs. true (3,787 validation materials)02460246true dir_gap (eV)predicted dir_gap heat_all: predicted from the structure vs. true (3,787 validation materials)024024true heat_all (eV/atom)predicted heat_all cosine similarity of the two latents, per material0.650.70.750.80.850200400600800cosine(structure latent, property latent)count0.636\u20130.642: 10.672\u20130.678: 10.689\u20130.695: 10.695\u20130.701: 10.701\u20130.707: 20.707\u20130.713: 30.713\u20130.719: 30.719\u20130.725: 60.725\u20130.731: 50.731\u20130.737: 160.737\u20130.743: 280.743\u20130.748: 520.748\u20130.754: 1080.754\u20130.76: 2970.76\u20130.766: 7360.766\u20130.772: 7540.772\u20130.778: 5820.778\u20130.784: 7710.784\u20130.79: 2510.79\u20130.796: 640.796\u20130.802: 290.802\u20130.808: 290.808\u20130.813: 180.813\u20130.819: 170.819\u20130.825: 50.825\u20130.831: 30.831\u20130.837: 30.855\u20130.861: 1mean"},{"location":"benchmarks/results/shipped-checkpoint-val-split.html#notes","title":"Notes","text":"<p>What this code measures on the shipped checkpoint, with every prediction saved (parity plots below). The paper's representation row quotes a cosine of about 0.97; this re-run does not reach it, see the comparison on that row's page.</p>"},{"location":"community/index.html","title":"Community results","text":"<p>Every submitted result, newest first. Contribute yours.</p> date result dataset category evidence status 2026-10-03 MLIP screening of the shipped paper candidates (MACE-MP-0) Perov-5 Validation level \u25cf\u25cf\u25cf\u25cb\u25cb MLIP validated Community submitted 2026-10-03 Shipped checkpoint re-evaluated on the training split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-10-03 Shipped checkpoint re-evaluated on the held-out validation split (this code) Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated MEIDNet verified 2026-05-29 MEIDNet (paper): inverse design of perovskites from property targets Perov-5 Conditional design \u25cf\u25cf\u25cf\u25cf\u25cb DFT validated Published (from the paper) 2026-05-29 MEIDNet (paper): alignment of structure and property latents on Perov-5 Perov-5 Representation quality \u25cf\u25cb\u25cb\u25cb\u25cb generated Published (from the paper)"},{"location":"community/contribute.html","title":"Contribute a benchmark result","text":"<p>MEIDNet Benchmarks collect results per dataset, each with the evidence behind it. Anyone who runs MEIDNet on a dataset \u2014 the published Perov-5, or their own \u2014 can submit a result. Submissions go through GitHub, so every number has a public record and a reviewer.</p>"},{"location":"community/contribute.html#what-a-submission-is","title":"What a submission is","text":"<p>One JSON file, <code>benchmarks/submissions/&lt;dataset&gt;/&lt;id&gt;.json</code>, that follows <code>benchmarks/schema.json</code>. The short version:</p> field what to put <code>dataset</code> an id from <code>benchmarks/datasets/</code> (<code>perov5</code>, \u2026); a new dataset needs its own <code>datasets/&lt;id&gt;.json</code> with source, size and split <code>modalities</code> what the model aligned: <code>structure</code>, <code>property:&lt;column&gt;</code> (one per property), later <code>xrd</code>, \u2026 <code>model</code> MEIDNet version, family and variant, the configuration (<code>meidnet.yaml</code>), checkpoint link and its SHA-256 <code>category</code> <code>representation_quality</code>, <code>conditional_design</code>, <code>efficiency</code> or <code>validation</code> \u2014 one category per file <code>results</code> metric \u2192 <code>{value, unit}</code>; only the metric names allowed for that category (listed in the schema) <code>validation_level</code> the evidence ladder: <code>generated</code> \u2192 <code>ml_filtered</code> \u2192 <code>mlip_validated</code> \u2192 <code>dft_validated</code> \u2192 <code>experimentally_validated</code> <code>evidence</code> public links: CIF archive, <code>stability.csv</code>, DFT inputs/outputs, paper DOI <code>reproduce</code> configuration, seed and the exact command that produces the numbers <code>status</code> always <code>community_submitted</code> for a new file"},{"location":"community/contribute.html#common-mistakes","title":"Common mistakes","text":"<ol> <li>Mixing categories in one file. A representation result (MAE, R\u00b2, retrieval) and a design result (SUN rate)    are two files.</li> <li>A metric name that is not in the list. <code>validate</code> refuses it; the lists exist so that rows are comparable.</li> <li>No split. State the split you evaluated on (or link a file of ids). A number on an unknown split is not comparable.</li> <li>Evidence that is not public. A result without an inspectable artefact stays at <code>generated</code>.</li> <li>Claiming a validation level the evidence does not support. <code>dft_validated</code> needs the DFT outputs, not the intention.</li> </ol>"},{"location":"community/contribute.html#how-it-works","title":"How it works","text":"<ol> <li>Copy <code>benchmarks/submissions/_template.json</code>, fill it in, run <code>python scripts/benchmarks.py validate</code>.</li> <li>Open a pull request (the template has a checklist). CI runs <code>validate</code>.</li> <li>Once merged, the result is rendered on the dataset's benchmark page and gets a permanent page of its own,    marked Community submitted.</li> <li>The maintainer may run <code>python scripts/benchmarks.py reproduce &lt;id&gt;</code>: it re-runs the stated command from the    stated configuration and compares the numbers. If they agree within the tolerances, the row becomes    MEIDNet verified (a record in <code>benchmarks/verified/</code>, which only the maintainer can add).</li> </ol> <p>Verified means reproduced from the submitted configuration with this code. It does not mean the material is real: that is what the validation level says. A re-run that does not agree is kept too (<code>benchmarks/verified/&lt;id&gt;.attempt.json</code>) and shown on the row's page next to the submitted numbers, so that disagreements are visible. This applies to the paper's own rows as well.</p> <p>Representation rows that are reproduced here also get their per-material predictions saved (<code>benchmarks/verified/&lt;id&gt;.predictions.csv</code>), which the pages turn into parity plots and a latent-agreement histogram.</p> <p>If you prefer not to open a pull request, the issue template Benchmark submission asks for the same fields and the maintainer converts it. Submissions are reviewed by the author of MEIDNet.</p>"},{"location":"examples/double-perovskite.html","title":"Double perovskites A\u2082BB\u2032X\u2086","text":"<p><code>meidnet/families/double_perovskite_a2bbx6.yaml</code> is a second family written with the same ingredients as the perovskite one. It shows what a family file can express:</p> <ul> <li>a 10-site prototype in the rhombohedral primitive cell of an fcc lattice (<code>lattice: fcc_primitive</code>:   edge a/\u221a2, angles 60\u00b0), A at (\u00bc,\u00bc,\u00bc) and (\u00be,\u00be,\u00be), B at the origin, B\u2032 at (\u00bd,\u00bd,\u00bd), six X;</li> <li>four site groups with their own element lists and oxidation states (B: +1/+2, B\u2032: +3/+4);</li> <li>a cell-size rule <code>a = 2 (r_B + r_B\u2032 + 2 r_X)</code> (one B\u2013X\u2013B\u2032 chain per half cell edge);</li> <li>rules that average radii over several groups: <code>tolerance_factor: {A: A, B: [B1, B2], X: X}</code>.</li> </ul> <pre><code>meidnet families double_perovskite_a2bbx6 --variant halide\n</code></pre> <pre><code>Rock-salt ordered double perovskite A2BB'X6 (Fm-3m)  [variant: halide]\n  formula pattern: A2B1B2X6   (10 atoms per cell)\n  A: 9 candidate elements - Large cation in the cuboctahedral cavity\n  B1: 15 candidate elements - First octahedral cation (usually +1 or +2)\n  B2: 20 candidate elements - Second octahedral cation (usually +3 or +4)\n  X: 4 candidate elements - Anion shared between neighbouring B and B' octahedra\n</code></pre> <p>The halide variant has 10,800 compositions; 305 pass every rule (Cs\u2082AgBiI\u2086 among them). Candidates are built in the 10-atom primitive cell; the symmetry refinement step writes them out as the conventional 40-atom Fm-3m cell, which is what the CIF files contain.</p>"},{"location":"examples/double-perovskite.html#using-it","title":"Using it","text":"<p>A model for this family must be trained on double-perovskite structures aligned to the prototype:</p> <pre><code>family: double_perovskite_a2bbx6\ndata:\n  table: double_perovskites.csv\n  properties: [{column: band_gap, unit: eV}]\n  align_to_prototype: true\ngeneration:\n  family: double_perovskite_a2bbx6\n  variant: halide\n  objectives: [{property: band_gap}]\n  targets: [{band_gap: 2.0}]\n</code></pre> <p>The published checkpoint</p> <p>The Perov-5 model was trained on 5-atom ABX\u2083 cells. The Studio will score double-perovskite compositions with it (the encoder accepts any cell up to 20 atoms), but those predictions are far outside its training distribution \u2014 a demonstration of the mechanics, not a result.</p>"},{"location":"examples/perov5.html","title":"Perov-5 \u2014 the paper's experiment","text":"<p>The published MEIDNet model: cubic ABX\u2083 perovskites from the Perov-5 dataset (CDVAE split of Castelli et al.), properties formation enthalpy (<code>heat_all</code>, eV/atom) and direct band gap (<code>dir_gap</code>, eV).</p>"},{"location":"examples/perov5.html#use-the-published-checkpoint","title":"Use the published checkpoint","text":"<pre><code>meidnet demo --family halide --band-gap 2.0            # or: oxide, chalcogenide, nitride\nmeidnet studio                                          # explore it live\n</code></pre> <p>The three checkpoints from the paper are in <code>checkpoints/</code>; <code>..._propertyaware_2k.pth</code> is the recommended one. <code>meidnet info checkpoints/&lt;file&gt;</code> describes any of them. They load into MEIDNet 2 unchanged.</p>"},{"location":"examples/perov5.html#retrain-from-scratch","title":"Retrain from scratch","text":"<pre><code>meidnet download-data                       # train/val/test CSVs \u2192 data/perov5/ (18,928 rows, 18 MB)\nmeidnet train examples/perov5/meidnet.yaml  # 200 epochs \u2248 1 h on a laptop GPU\nmeidnet generate examples/perov5/meidnet.yaml\n</code></pre> <p><code>examples/perov5/meidnet.yaml</code> reproduces the paper's settings, including two that differ from MEIDNet 2's defaults: <code>normalize: false</code> for both properties and <code>align_to_prototype: false</code> (atoms in file order), and the 1200-epoch alignment ramp.</p>"},{"location":"examples/perov5.html#the-papers-generated-structures","title":"The paper's generated structures","text":"<p><code>examples/perov5/paper_results/</code> holds the 26 CIFs (halide, oxide, chalcogenide) reported in the paper. <code>meidnet screen examples/perov5/paper_results --train-csv data/perov5/train.csv</code> recomputes their MACE stability.</p>"},{"location":"examples/perov5.html#regression-against-v1","title":"Regression against v1","text":"<p><code>tests/legacy_v1/</code> contains the v1.0 code byte for byte. <code>pytest -m slow</code> runs it and MEIDNet 2 on the same targets and asserts identical CIFs, predictions and file names.</p>"},{"location":"explore/capabilities.html","title":"Capabilities","text":"<p>This page is generated from the code. Available: in this release. Community-tested: used by a benchmark result reproduced here. Planned: on the roadmap, not yet implemented.</p>"},{"location":"explore/capabilities.html#modalities","title":"Modalities","text":"modality status how Crystal structure (CIF, up to <code>max_sites</code> atoms, default 20) Community-tested an equivariant graph encoder; structures are aligned to the family prototype Scalar properties (any number of numeric columns) Community-tested one property encoder; every column becomes a target you can set Published properties: direct band gap (<code>dir_gap</code>), formation enthalpy (<code>heat_all</code>) Community-tested the shipped Perov-5 checkpoint Vector modalities: binned XRD, DOS Planned an encoder per vector modality into the shared latent space Spectra (Raman, UV-Vis) Planned as vector modalities Text (descriptions, synthesis) Planned a text encoder into the shared latent space Images (microscopy) Planned an image encoder into the shared latent space"},{"location":"explore/capabilities.html#fusion","title":"Fusion","text":"<p>The structure latent and the property latent of a material are averaged into the joint latent (<code>meidnet/model.py</code>); the paper calls this early fusion. The two latents are first aligned by a contrastive loss whose weight increases over <code>training.contrastive_warmup_epochs</code> (the curriculum of the paper). Other fusion schemes are compared in the Architecture Atlas.</p>"},{"location":"explore/capabilities.html#input-formats","title":"Input formats","text":"format note CSV any delimiter pandas reads Excel (.xlsx, .xls) first sheet JSON table records Parquet needs the optional pyarrow CIF text in a column one structure per row CIF files (.cif, or a .zip of them) named .cif"},{"location":"explore/capabilities.html#material-families","title":"Material families","text":"family variants site groups <code>double_perovskite_a2bbx6</code> - Rock-salt ordered double perovskite A2BB'X6 (Fm-3m) <code>halide</code>, <code>oxide</code> A, B1, B2, X <code>perovskite_abx3</code> - Cubic ABX3 perovskite (Pm-3m) <code>oxide</code>, <code>halide</code>, <code>chalcogenide</code>, <code>nitride</code> A, B, X <p>Any other prototype family can be described in a family file.</p>"},{"location":"explore/capabilities.html#rules-hard-constraints","title":"Rules (hard constraints)","text":"<ul> <li><code>bond_window</code> - Sensible {from}\u2013{to} bonds</li> <li><code>charge_neutrality</code> - Charge balance</li> <li><code>min_distance</code> - No overlapping atoms</li> <li><code>octahedral_factor</code> - Octahedral factor</li> <li><code>property_window</code> - Predicted {property} window</li> <li><code>tolerance_factor</code> - Goldschmidt tolerance factor</li> </ul>"},{"location":"explore/capabilities.html#search-terms-soft","title":"Search terms (soft)","text":"<ul> <li><code>cubic_cell</code> - Pulls the decoded cell towards a cube of the family's reference volume.</li> <li><code>entropy_bonus</code> - Rewards uncertainty early in a round so the search explores (fades to zero).</li> <li><code>group_consistency</code> - Sites of the same group should agree on their element (ramps up during a round).</li> <li><code>prototype_alignment</code> - Pulls decoded positions onto the prototype's site positions.</li> <li><code>short_distance_penalty</code> - Large penalty if any two decoded atoms come closer than <code>min</code>.</li> <li><code>site_repulsion</code> - Short-range exponential repulsion between decoded atoms.</li> <li><code>site_separation</code> - Keeps decoded atoms from collapsing onto each other (1/distance).</li> <li><code>soft_charge</code> - Expected total charge of the most likely composition should be zero (ramps up during a round).</li> <li><code>tolerance_penalty</code> - Soft version of the Goldschmidt tolerance-factor window, using the most likely elements.</li> </ul>"},{"location":"explore/capabilities.html#after-generation","title":"After generation","text":"<ul> <li><code>meidnet screen</code>: MACE-MP-0 relaxation and the stable / unique / novel (SUN) counts (optional <code>pip install meidnet[stability]</code>). DFT and experiments are yours: the benchmark evidence ladder records them.</li> </ul>"},{"location":"explore/databases.html","title":"Databases by application","text":"<p>Where to find training data and realistic targets: computed (DFT) databases give structures with properties, experimental databases give measured values and refined structures. MEIDNet needs, per material, one crystal structure plus one or more scalar properties (what data do I need?); the last column says how far each source is from that.</p> <p>Pick an application:</p> <ul> <li>Solar cells and photovoltaics</li> <li>Semiconductor physics</li> <li>Batteries and ionic conductors</li> <li>Thermoelectrics</li> <li>Catalysis and surfaces</li> <li>Magnetism and superconductivity</li> <li>Mechanical, dielectric and piezoelectric</li> <li>2D materials</li> <li>Porous materials and MOFs</li> <li>Generative-model benchmarks</li> <li>All databases</li> </ul>"},{"location":"explore/databases.html#solar-cells-and-photovoltaics","title":"Solar cells and photovoltaics","text":"<p>What to look for: band gap in the 1.1-1.8 eV window (single junction) or 1.6-2.0 eV (tandem top cell), direct gap, strong absorption, formation energy / hull distance, defect tolerance, device records for the experimental side</p> <p>As MEIDNet targets: dir_gap or any gap column + heat_all / e_above_hull; family = perovskite_abx3 (halide, oxide, chalcogenide), or your own prototype</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure NOMAD computed more than 10,000,000 calculations raw and normalised DFT outputs from many codes; band structures, DOS, energies; FAIR provenance needs curation: pick one code and functional before building a table C2DB computed about 4,000 monolayers stability (dynamic and thermodynamic), PBE / HSE / GW gaps, magnetic state, optical absorption, piezoelectric and Raman data for monolayers ready for a 2D prototype family Cubic perovskites (CMR, Castelli et al.) computed about 19,000 ABX3 compositions formation (heat of formation) energies and direct / indirect gaps (GLLB-SC) for cubic ABX3 with O, N, S, F anions and their mixtures - the source of Perov-5 the published MEIDNet benchmark (Perov-5 split): meidnet download-data matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Perovskite Database Project experimental more than 42,000 solar-cell devices device-level records: composition, architecture, efficiency, Voc, Jsc, FF, stability, processing conditions, from the literature no crystal structures per record: use it to choose realistic gap / composition targets and to check candidates against what has been made Hybrid organic-inorganic perovskites (Kim et al. 2017) computed 1,346 HOIPs DFT structures, band gaps and dielectric constants of ABX3 hybrid perovskites ready, but organic A-site cations need a family file with molecular site groups (not shipped) Experimental band gaps (Zhuo et al. 2018) experimental 6,354 compounds measured band gaps by composition, with the matbench_expt_gap subset as a standard task targets and validation of predicted gaps; no structures Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#semiconductor-physics","title":"Semiconductor physics","text":"<p>What to look for: band gaps at several levels of theory (PBE, HSE, mBJ) and experimental, effective masses, dielectric constants, carrier mobility proxies, band edges, phonons</p> <p>As MEIDNet targets: gap, effective mass or dielectric constant as scalar targets; keep the level of theory consistent inside one table</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) NOMAD computed more than 10,000,000 calculations raw and normalised DFT outputs from many codes; band structures, DOS, energies; FAIR provenance needs curation: pick one code and functional before building a table Materials Cloud computed curated archives (MC3D, MC2D, phonons, ...) MC3D relaxed structures, MC2D exfoliable monolayers, phonon database, Sssp pseudopotential sets, workflow provenance (AiiDA) ready for MC3D / MC2D tables C2DB computed about 4,000 monolayers stability (dynamic and thermodynamic), PBE / HSE / GW gaps, magnetic state, optical absorption, piezoelectric and Raman data for monolayers ready for a 2D prototype family Cubic perovskites (CMR, Castelli et al.) computed about 19,000 ABX3 compositions formation (heat of formation) energies and direct / indirect gaps (GLLB-SC) for cubic ABX3 with O, N, S, F anions and their mixtures - the source of Perov-5 the published MEIDNet benchmark (Perov-5 split): meidnet download-data Matbench and Matbench Discovery computed 13 tasks (Matbench); 257,000 WBM test structures (Discovery) standardised property-prediction tasks (gap, formation energy, moduli, dielectric, ...) with fixed folds; Discovery ranks universal potentials on stability prediction with F1, DAF, RMSD and a leaderboard with plots use the tasks as property tables; the Discovery leaderboard is the model for MEIDNet Benchmarks matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Hybrid organic-inorganic perovskites (Kim et al. 2017) computed 1,346 HOIPs DFT structures, band gaps and dielectric constants of ABX3 hybrid perovskites ready, but organic A-site cations need a family file with molecular site groups (not shipped) Experimental band gaps (Zhuo et al. 2018) experimental 6,354 compounds measured band gaps by composition, with the matbench_expt_gap subset as a standard task targets and validation of predicted gaps; no structures QMOF computed about 20,000 MOFs DFT-optimised MOF structures with band gaps and charges beyond today's cell-size limit; targets and descriptors Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#batteries-and-ionic-conductors","title":"Batteries and ionic conductors","text":"<p>What to look for: ionic conductivity, migration barriers, voltage, stability window, volume change, hull distance</p> <p>As MEIDNet targets: formation energy + conductivity (log scale) as targets; families with a mobile-ion site group</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Liverpool Ionic Conductivity Database (LiIonDB) experimental about 800 measurements experimental Li-ion conductivities with temperature and the reported phase, from the literature targets (log conductivity); join with COD / Materials Project structures Materials Project battery explorer computed about 4,000 intercalation electrodes average voltage, capacity, volume change and stability of intercalation electrodes computed from the MP structures ready: voltage and capacity as targets on the host-structure family Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#thermoelectrics","title":"Thermoelectrics","text":"<p>What to look for: Seebeck coefficient, electrical and thermal conductivity, zT, carrier concentration and temperature; experimental data is temperature-resolved</p> <p>As MEIDNet targets: zT or Seebeck at a fixed temperature as a scalar target</p> database kind size features use with MEIDNet AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets UCSB thermoelectrics (Gaultois et al.) experimental about 1,100 compounds experimental Seebeck, resistivity, thermal conductivity, zT at a given temperature targets; join with structures Starrydata experimental tens of thousands of digitised curves temperature-dependent thermoelectric and other property curves digitised from papers targets at a chosen temperature Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#catalysis-and-surfaces","title":"Catalysis and surfaces","text":"<p>What to look for: adsorption energies, reaction barriers, surface energies, work functions</p> <p>As MEIDNet targets: bulk descriptors (formation energy, d-band proxies) only - MEIDNet works on bulk prototypes, not slabs</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table NOMAD computed more than 10,000,000 calculations raw and normalised DFT outputs from many codes; band structures, DOS, energies; FAIR provenance needs curation: pick one code and functional before building a table Open Catalyst (OC20 / OC22) computed more than 1,300,000 relaxations adsorbate-surface relaxations with energies and forces; the standard catalysis ML benchmark surfaces are out of MEIDNet's scope; use the bulk subsets as structure sources Catalysis-Hub computed more than 100,000 reaction energies adsorption and reaction energies on surfaces with the DFT settings used descriptors only Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#magnetism-and-superconductivity","title":"Magnetism and superconductivity","text":"<p>What to look for: magnetic moment and ordering, Curie / Neel temperature, superconducting Tc with the structure it belongs to</p> <p>As MEIDNet targets: total magnetization per formula unit, Tc (log scale) as scalar targets</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally SuperCon (NIMS) and 3DSC experimental about 16,000 Tc entries; 3DSC links 5,700 to structures experimental superconducting critical temperatures; 3DSC matches them to ICSD / Materials Project structures 3DSC is ready (structure + Tc); use log Tc as the target MAGNDATA (Bilbao) experimental about 2,000 magnetic structures experimentally determined commensurate and incommensurate magnetic structures structures with magnetic ordering; join with moments as targets"},{"location":"explore/databases.html#mechanical-dielectric-and-piezoelectric","title":"Mechanical, dielectric and piezoelectric","text":"<p>What to look for: bulk and shear moduli, hardness, dielectric tensor, piezoelectric coefficients, phonon stability</p> <p>As MEIDNet targets: bulk modulus, dielectric constant or a piezoelectric scalar as targets</p> database kind size features use with MEIDNet Materials Project computed about 150,000 inorganic compounds relaxed structures, formation energy, energy above hull, PBE/r2SCAN band gaps, magnetic ordering, elastic / dielectric / piezoelectric tensors, phonons for a subset, X-ray absorption, battery (intercalation) electrodes ready: export CIFs + a property table with the API; one prototype family per table AFLOW computed more than 3,500,000 entries formation enthalpy, band gaps, elastic and thermal properties (AGL), Debye temperature, magnetic moments, prototype encyclopedia ready; the AFLOW prototype library is a good source of family files JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Materials Cloud computed curated archives (MC3D, MC2D, phonons, ...) MC3D relaxed structures, MC2D exfoliable monolayers, phonon database, Sssp pseudopotential sets, workflow provenance (AiiDA) ready for MC3D / MC2D tables Matbench and Matbench Discovery computed 13 tasks (Matbench); 257,000 WBM test structures (Discovery) standardised property-prediction tasks (gap, formation energy, moduli, dielectric, ...) with fixed folds; Discovery ranks universal potentials on stability prediction with F1, DAF, RMSD and a leaderboard with plots use the tasks as property tables; the Discovery leaderboard is the model for MEIDNet Benchmarks matminer datasets computed + experimental about 50 curated tables one-line loaders for experimental band gaps (Zhuo 2018, 6,354 compounds; matbench_expt_gap 4,604), experimental formation enthalpies (Kim 2017), UCSB thermoelectrics, elastic tensors, dielectric constants, piezoelectric tensors, phonon data, superhard materials, HOIP perovskites experimental tables are composition-only: join with a structure source (COD, Materials Project) to train; use them directly as realistic targets Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files ICSD experimental about 300,000 inorganic structures the reference collection of experimentally determined inorganic structures; the ground truth most DFT databases start from structures: licence forbids redistribution, train locally Materials Data Facility computed + experimental hundreds of datasets a registry of published materials datasets, experimental and computed, with DOIs a place to find and to publish tables"},{"location":"explore/databases.html#2d-materials","title":"2D materials","text":"<p>What to look for: exfoliation energy, gap, magnetic state, dynamic stability of monolayers</p> <p>As MEIDNet targets: a 2D prototype family (e.g. MXene or TMD) with gap and exfoliation energy as targets</p> database kind size features use with MEIDNet JARVIS-DFT (NIST) computed about 80,000 3D and 1,000 2D materials OptB88vdW and TBmBJ band gaps, effective masses, dielectric functions, solar-cell efficiency (SLME), elastic tensors, piezoelectric and thermoelectric (BoltzTraP) properties, exfoliation energies, superconducting Tc (electron-phonon) for a subset ready; the richest single source of scalar targets per structure Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) Materials Cloud computed curated archives (MC3D, MC2D, phonons, ...) MC3D relaxed structures, MC2D exfoliable monolayers, phonon database, Sssp pseudopotential sets, workflow provenance (AiiDA) ready for MC3D / MC2D tables C2DB computed about 4,000 monolayers stability (dynamic and thermodynamic), PBE / HSE / GW gaps, magnetic state, optical absorption, piezoelectric and Raman data for monolayers ready for a 2D prototype family 2DMatPedia computed about 6,000 monolayers exfoliation energy, band gap, magnetic moment of monolayers derived from bulk databases ready for a 2D prototype family"},{"location":"explore/databases.html#porous-materials-and-mofs","title":"Porous materials and MOFs","text":"<p>What to look for: pore geometry, gas uptake, band gap of the framework, stability</p> <p>As MEIDNet targets: frameworks exceed MEIDNet's cell-size limit today (max_sites); use the property tables to set targets and the descriptors as inspiration</p> database kind size features use with MEIDNet Crystallography Open Database (COD) experimental more than 500,000 crystal structures experimentally determined structures (organic and inorganic) as CIFs with references; no properties structures: join with a property table to train; the natural source of experimental prototypes for family files Cambridge Structural Database (CSD) experimental more than 1,300,000 organic and metal-organic structures experimental molecular and MOF crystal structures beyond today's cell-size limit; descriptors only QMOF computed about 20,000 MOFs DFT-optimised MOF structures with band gaps and charges beyond today's cell-size limit; targets and descriptors CoRE MOF experimental about 14,000 experimental MOFs computation-ready experimental MOF structures with pore descriptors beyond today's cell-size limit"},{"location":"explore/databases.html#generative-model-benchmarks","title":"Generative-model benchmarks","text":"<p>What to look for: fixed splits used by CDVAE, DiffCSP, MatterGen and MEIDNet; validity, uniqueness, novelty and stability metrics</p> <p>As MEIDNet targets: the Perov-5 split is the published MEIDNet benchmark; MP-20 and Carbon-24 test structure representation</p> database kind size features use with MEIDNet OQMD computed more than 1,000,000 structures formation energy, stability (hull), band gap (PBE), many hypothetical prototypes (Heusler, perovskite, ...) ready: prototype-decorated entries are exactly what a MEIDNet family describes Alexandria computed about 4,500,000 PBE and 400,000 PBEsol / SCAN structures formation energy, hull distance, band gap, magnetic moment, for a very large set of hypothetical compounds (1D, 2D, 3D) ready; the usual pre-training set of generative models (MatterGen's Alex-MP-20 split) GNoME (DeepMind) computed about 380,000 stable crystals structures predicted stable with DFT energies and hull distances; no electronic properties structures only: join with a computed property before training Cubic perovskites (CMR, Castelli et al.) computed about 19,000 ABX3 compositions formation (heat of formation) energies and direct / indirect gaps (GLLB-SC) for cubic ABX3 with O, N, S, F anions and their mixtures - the source of Perov-5 the published MEIDNet benchmark (Perov-5 split): meidnet download-data CDVAE splits (Perov-5, MP-20, Carbon-24) computed 18,928 / 45,231 / 10,153 structures fixed train / val / test splits used by generative-model papers; Perov-5 carries formation energy and band gap, the others energies only ready: the format meidnet reads directly (id, cif, property columns) Matbench and Matbench Discovery computed 13 tasks (Matbench); 257,000 WBM test structures (Discovery) standardised property-prediction tasks (gap, formation energy, moduli, dielectric, ...) with fixed folds; Discovery ranks universal potentials on stability prediction with F1, DAF, RMSD and a leaderboard with plots use the tasks as property tables; the Discovery leaderboard is the model for MEIDNet Benchmarks"},{"location":"explore/databases.html#all-databases","title":"All databases","text":"database kind access licence applications Materials Project computed web, mp-api (free key) CC BY 4.0 Solar cells and photovoltaics, Semiconductor physics, Batteries and ionic conductors, Magnetism and superconductivity, Mechanical, dielectric and piezoelectric, Catalysis and surfaces OQMD computed web, REST API, full download open Solar cells and photovoltaics, Semiconductor physics, Magnetism and superconductivity, Generative-model benchmarks AFLOW computed web, REST / AFLUX API open Semiconductor physics, Mechanical, dielectric and piezoelectric, Thermoelectrics, Magnetism and superconductivity JARVIS-DFT (NIST) computed web, jarvis-tools (figshare downloads) open (NIST) Solar cells and photovoltaics, Semiconductor physics, Thermoelectrics, Mechanical, dielectric and piezoelectric, 2D materials, Magnetism and superconductivity Alexandria computed bulk download CC BY 4.0 Generative-model benchmarks, Semiconductor physics, Magnetism and superconductivity, 2D materials GNoME (DeepMind) computed download CC BY-NC 4.0 Generative-model benchmarks NOMAD computed web, API, raw calculation files CC BY 4.0 Semiconductor physics, Solar cells and photovoltaics, Catalysis and surfaces Materials Cloud computed web, archive downloads CC BY 4.0 (per archive) 2D materials, Mechanical, dielectric and piezoelectric, Semiconductor physics C2DB computed web, ASE database download CC BY 4.0 2D materials, Semiconductor physics, Solar cells and photovoltaics 2DMatPedia computed web, download open 2D materials Cubic perovskites (CMR, Castelli et al.) computed web, ASE database download CC BY 4.0 Solar cells and photovoltaics, Semiconductor physics, Generative-model benchmarks CDVAE splits (Perov-5, MP-20, Carbon-24) computed download (CSV with CIF column) MIT (repository) Generative-model benchmarks Matbench and Matbench Discovery computed web, matbench / matbench-discovery packages open Generative-model benchmarks, Semiconductor physics, Mechanical, dielectric and piezoelectric matminer datasets computed + experimental matminer.datasets.load_dataset per dataset Semiconductor physics, Solar cells and photovoltaics, Thermoelectrics, Mechanical, dielectric and piezoelectric Crystallography Open Database (COD) experimental web, full download, REST public domain (CC0) Semiconductor physics, Solar cells and photovoltaics, Batteries and ionic conductors, Magnetism and superconductivity, Mechanical, dielectric and piezoelectric, Porous materials and MOFs ICSD experimental licence (most universities have it) commercial Semiconductor physics, Solar cells and photovoltaics, Batteries and ionic conductors, Magnetism and superconductivity, Mechanical, dielectric and piezoelectric Cambridge Structural Database (CSD) experimental licence commercial (CSD-Community subset free) Porous materials and MOFs Perovskite Database Project experimental web, download CC BY 4.0 Solar cells and photovoltaics Hybrid organic-inorganic perovskites (Kim et al. 2017) computed download (Scientific Data), matminer CC BY 4.0 Solar cells and photovoltaics, Semiconductor physics Experimental band gaps (Zhuo et al. 2018) experimental paper SI, matminer expt_gap per paper Semiconductor physics, Solar cells and photovoltaics Liverpool Ionic Conductivity Database (LiIonDB) experimental web, download open Batteries and ionic conductors Materials Project battery explorer computed web, mp-api CC BY 4.0 Batteries and ionic conductors UCSB thermoelectrics (Gaultois et al.) experimental matminer ucsb_thermoelectrics open Thermoelectrics Starrydata experimental web, download open Thermoelectrics Open Catalyst (OC20 / OC22) computed download, fairchem CC BY 4.0 Catalysis and surfaces Catalysis-Hub computed web, GraphQL API open Catalysis and surfaces SuperCon (NIMS) and 3DSC experimental NIMS MDR, 3DSC on GitHub open (3DSC) Magnetism and superconductivity MAGNDATA (Bilbao) experimental web open Magnetism and superconductivity QMOF computed download (figshare) CC BY 4.0 Porous materials and MOFs, Semiconductor physics CoRE MOF experimental download CC BY 4.0 Porous materials and MOFs Materials Data Facility computed + experimental web, API per dataset Semiconductor physics, Solar cells and photovoltaics, Batteries and ionic conductors, Thermoelectrics, Catalysis and surfaces, Mechanical, dielectric and piezoelectric <p>Missing one you use? Open an issue or edit <code>benchmarks/catalog/databases.json</code>.</p>"},{"location":"explore/datasets.html","title":"Datasets","text":"<p>Datasets used with MEIDNet and what has been done with each. Tasks, sizes and properties differ from one dataset to the next, so results are compared within a dataset only (see Benchmarks).</p> dataset structures properties used size available here Perov-5 cubic ABX\u2083 perovskites, 5 atoms per cell direct band gap, formation enthalpy 18,928 pretrained model, configuration, tutorial, benchmark results MP-20 Materials Project, up to 20 atoms none aligned in the paper about 45,000 structure-representation results in the paper Carbon-24 carbon allotropes, up to 24 atoms none aligned in the paper about 10,000 structure-representation results in the paper Double perovskites A\u2082BB\u2032X\u2086 a family file, no dataset none 10,800 compositions (halide variant) design space and rules Your dataset one material family any scalar columns any upload in the Studio, or the command line <p>New to multimodal data? Learn multimodality explains what a modality is, and Databases by application lists 31 computed and experimental sources.</p> <ul> <li> <p>Perov-5 \u2014 published multimodal benchmark</p> <p>18,928 cubic ABX\u2083 perovskites (CDVAE split: 11,356 train / 3,787 validation / 3,785 test), each with a relaxed structure, formation enthalpy (<code>heat_all</code>, eV/atom) and direct band gap (<code>dir_gap</code>, eV).</p> <p>Modalities used in MEIDNet: crystal structure \u00b7 electronic property (band gap) \u00b7 thermodynamic property (formation enthalpy).</p> <p>What the paper does with it: multimodal alignment of the structure and property encoders, property reconstruction, and the inverse-design demonstration (candidates generated from property targets, screened with MACE and validated by DFT).</p> <p>In this repository: <code>meidnet download-data</code> fetches it; the published checkpoint <code>checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth</code> was trained on it; it is the data behind the live Studio.</p> <p>Suitable tasks: inverse design from a band gap and a stability target \u00b7 property prediction from the structure \u00b7 structure\u2013property retrieval \u00b7 measuring how well two modalities align.</p> <p>Available here: pretrained model \u00b7 configuration (<code>examples/perov5/meidnet.yaml</code>) \u00b7 tutorial (the paper's experiment, Colab) \u00b7 benchmark results.</p> <p>Open in the Studio</p> <p>Benchmark results \u00b7 The paper's experiment \u00b7 Dataset source (CDVAE) \u00b7 Castelli et al. 2012, Xie et al. 2022</p> </li> <li> <p>MP-20 \u2014 published: structure-representation generalization</p> <p>~45,000 structures from the Materials Project with at most 20 atoms per cell (CDVAE split).</p> <p>Published MEIDNet use: generalization of the crystal encoder / decoder beyond perovskites (structure representation and reconstruction). It is not a multimodal benchmark in the paper: no property modality was aligned on it, and no MP-20 checkpoint is shipped here.</p> <p>Benchmark results \u00b7 Dataset source (CDVAE)</p> </li> <li> <p>Carbon-24 \u2014 published: structure-representation generalization</p> <p>~10,000 carbon structures with up to 24 atoms per cell (CDVAE split).</p> <p>Published MEIDNet use: the same structure-representation test as MP-20. Not a multimodal benchmark; no Carbon-24 checkpoint is shipped here.</p> <p>Benchmark results \u00b7 Dataset source (CDVAE)</p> </li> </ul>"},{"location":"explore/datasets.html#your-dataset","title":"Your dataset","text":"<p>Use your dataset in the Studio</p> <p>Any table with an id, one or more scalar property columns and a crystal structure per row (CIF text in a column or one <code>.cif</code> file per id) can be used directly: upload it in the Studio or run <code>meidnet init</code> \u2192 <code>check</code> \u2192 <code>train</code> \u2192 <code>generate</code> (bring your own dataset). The structures must belong to one material family (prototype + site groups), which is what the rules and the search need.</p>"},{"location":"explore/datasets.html#next-datasets","title":"Next datasets","text":"<p>Battery conductors, MOFs, 2D materials and spinels are natural next families. Results on any of them can be contributed: the dataset gets its own benchmark page, and a result reproduced here is marked as verified.</p>"},{"location":"learn/index.html","title":"Learn multimodality","text":"<p>A material can be described in many ways: by its atoms and their positions, by a formula, by measured or computed properties, by a diffraction pattern, a spectrum, a micrograph or a paragraph in a paper. Each way of describing it is a modality. Multimodal learning trains one model on several modalities of the same materials, so that it can connect them: read a structure and predict a property, or start from the properties you want and propose a structure, which is what MEIDNet does.</p> <p>This short course takes about ten minutes. It covers the main ideas of multimodal learning for materials and, for each, how MEIDNet implements it. For a full review, see Babu and Krishnan (2026).</p> <ol> <li>What is a modality?</li> <li>The multimodality map (interactive)</li> <li>Five challenges of multimodal learning</li> <li>Contrastive learning: how two modalities are aligned</li> <li>From a shared space to inverse design</li> <li>Where to go next</li> </ol>"},{"location":"learn/index.html#what-is-a-modality","title":"1. What is a modality?","text":"<p>A modality is one kind of observation of a material, with its own shape of data. The shape decides which kind of neural network can read it.</p> modality what the data looks like example a network that reads it Crystal structure atoms with positions in a unit cell: a graph a CIF file graph neural network Composition elements and their fractions: a set CsPbI\u2083 set or element network Scalar properties a short vector of numbers band gap, formation enthalpy small fully connected network (MLP) Diffraction pattern a curve: intensity versus angle 2\u03b8 powder XRD 1D convolutional network Spectra and DOS a curve: signal versus energy density of states, XAS, Raman 1D convolutional network or transformer Image a grid of pixels SEM or TEM micrograph 2D convolutional network or vision transformer Text a sequence of words a synthesis procedure language model <p>Why combine them? Each modality sees part of the material. A structure fixes the atoms but does not tell you the band gap without a calculation; a band gap says nothing about which atoms produce it. A model that learns both together can translate between them in either direction.</p> <p>What it needs: paired data, where several modalities describe the same material. In Perov-5, every one of the 18,928 structures comes with its DFT band gap and formation enthalpy, which makes it a two-modality dataset.</p>"},{"location":"learn/index.html#the-multimodality-map","title":"2. The multimodality map","text":"<p>Select a modality to see what it contains, which networks encode it, how it is combined with other modalities and how MEIDNet handles it. The centre describes the shared representation.</p> The map as a table modality MEIDNet how Crystal structure available an E(n)-equivariant graph network (EGNN) encodes up to <code>max_sites</code> atoms (default 20) Scalar properties available all numeric columns form one vector, read by one MLP; every column becomes a target Composition via the structure the element on each site is part of the structure; formula-only data needs a composition model Processing conditions partial numeric conditions can be extra scalar columns, encoded together with the properties Diffraction (XRD) planned a vector-modality encoder for binned patterns, first item of the roadmap Spectra and DOS planned as vector modalities Microscopy images planned an image encoder into the shared space Text planned a text encoder into the shared space <p>The capabilities page lists the same information, generated from the code.</p>"},{"location":"learn/index.html#five-challenges","title":"3. Five challenges of multimodal learning","text":"<p>Baltru\u0161aitis, Ahuja and Morency organise multimodal machine learning around five challenges. Every multimodal model answers each of them in some way, even if only by leaving it out.</p> challenge the question a materials example in MEIDNet Representation How is each modality encoded, and do they share one space (joint) or keep separate spaces tied by a constraint (coordinated)? a graph network for the structure, an MLP for the property vector both: two encoders whose outputs are coordinated by contrastive learning and then averaged into one joint latent of 128 numbers Translation How is one modality mapped to another? structure \u2192 properties (prediction), properties \u2192 structure (inverse design) both directions: decoders read the joint latent and the property latent alone Alignment Which parts of one modality correspond to which parts of another? which XRD peaks come from which lattice planes only whole materials: each structure is matched with its own property vector, not part by part Fusion How are the modalities joined to make a prediction? one descriptor from structure and band gap together the structure latent and the property latent are averaged (a note on the word \"early\") Co-learning Can one modality help learn another, for example a scarce one? learn from many simulated XRD patterns to read a few measured ones closest analogue: the property branch is trained to rebuild the crystal on its own, so at design time properties alone are enough <p>Further reading: Baltru\u0161aitis et al. 2019 (the taxonomy) \u00b7 Guo et al. 2019 (deep multimodal representation learning) \u00b7 Ngiam et al. 2011 (shared representations between modalities) \u00b7 Babu and Krishnan 2026 (multimodal and cross-modal learning techniques).</p>"},{"location":"learn/index.html#contrastive-learning","title":"4. Contrastive learning: how two modalities are aligned","text":"<p>Take a batch of \\(N\\) materials. Each gives two vectors of length one: \\(z_c\\) from its structure and \\(z_p\\) from its properties. The dot product of two such vectors is their cosine similarity. Lay out all \\(N \\times N\\) similarities in a table: the diagonal holds the true pairs, everything else is a structure next to someone else's properties.</p> <p>Contrastive learning turns each row into a multiple-choice question: which of the \\(N\\) property vectors belongs to this structure? The right answer is on the diagonal. It asks the same question for each column, from properties to structures. The loss MEIDNet uses is this symmetric InfoNCE loss:</p> \\[ \\mathcal{L}_\\text{InfoNCE} = \\frac{1}{2N}\\sum_{i=1}^{N}\\left[ -\\log\\frac{\\exp\\!\\big(z_c^{(i)}\\!\\cdot z_p^{(i)}/\\tau\\big)}{\\sum_{j=1}^{N}\\exp\\!\\big(z_c^{(i)}\\!\\cdot z_p^{(j)}/\\tau\\big)} -\\log\\frac{\\exp\\!\\big(z_p^{(i)}\\!\\cdot z_c^{(i)}/\\tau\\big)}{\\sum_{j=1}^{N}\\exp\\!\\big(z_p^{(i)}\\!\\cdot z_c^{(j)}/\\tau\\big)} \\right] \\] <p>The temperature \\(\\tau\\) sets how sharply the model must prefer the right answer. Move the slider to see its effect on a batch of five materials: each row shows how much probability the model gives to each candidate.</p> setting in MEIDNet value in <code>meidnet.yaml</code> temperature \\(\\tau\\) 0.01 <code>training.temperature</code> strength of the alignment 5.0 <code>training.contrastive_weight</code> warm-up: the strength grows from 0 to its full value over 60 % of the epochs (the paper: 1,200 epochs) <code>training.contrastive_warmup_epochs</code> size of the shared space 128 <code>model.latent_dim</code> <p>The warm-up is the paper's curriculum: the decoders first learn to rebuild crystals and properties, then the two latents are pulled together. After training, the training report measures the alignment with retrieval: how often the nearest property vector to a structure is its own (top 1) or among the five nearest (top 5). The Studio's Model block plots the mean cosine between matched pairs during training.</p> <p>The same loss pairs pictures with their captions in CLIP; MEIDNet pairs crystals with their properties.</p> <p>Further reading: van den Oord et al. 2018 (InfoNCE) \u00b7 Radford et al. 2021 (CLIP) \u00b7 Babu and Krishnan 2026 \u00b7 Contrastive learning in the Architecture Atlas.</p>"},{"location":"learn/index.html#inverse-design","title":"5. From a shared space to inverse design","text":"<p>Once both modalities live in one space, a point in it can be read in both directions. MEIDNet makes the reverse direction possible with one extra training term: the crystal decoder is also trained from the property latent alone. A vector of wanted properties can then be encoded and decoded into a structure.</p> <p>The search in practice:</p> <ol> <li>Encode the target. The property encoder turns the target values into a starting point in the shared space.</li> <li>Optimise a population of latent points so that their decoded properties approach the target, while staying    diverse and near the target's latent.</li> <li>Decode one element per site of a material family (a prototype such as cubic ABX\u2083), so candidates are    compositions placed on a known arrangement of sites.</li> <li>Check the rules (charge balance, tolerance factor, distances and your own) and keep only what passes.</li> <li>Rank by distance to the target and save each candidate as a CIF with a report.</li> </ol> <p>The full objective and every step are in How MEIDNet works. Two limits matter: MEIDNet does not invent new atomic arrangements, and its predicted properties are estimates to confirm with DFT or experiment (scope and roadmap).</p> <p>Further reading: the MEIDNet paper (Babu et al. 2026) \u00b7 Babu, Gouv\u00eaa and Rignanese 2026 (generative models, multimodal learning and closed-loop workflows in inverse materials design) \u00b7 Moro, Loh et al. 2025 (multimodal foundation models for materials).</p>"},{"location":"learn/index.html#where-next","title":"6. Where to go next","text":"<ul> <li> <p>Architecture Atlas</p> <p>Early fusion, late fusion, shared latent spaces, cross-attention and contrastive learning, drawn the same way, with an advisor that recommends one for your data.</p> <p> Architecture Atlas</p> </li> <li> <p>Recipes by problem</p> <p>From a research question to an architecture, a dataset, a configuration and a tutorial.</p> <p> Recipes by problem</p> </li> <li> <p>Build it</p> <p>Upload a table of structures and properties, train in the browser, set targets, run the search.</p> <p> Open the Studio</p> </li> <li> <p>Find data</p> <p>Datasets that work with MEIDNet, and 31 computed and experimental databases grouped by application.</p> <p> Datasets</p> </li> </ul>"},{"location":"learn/index.html#references","title":"References","text":"<ul> <li>T. Baltru\u0161aitis, C. Ahuja, L.-P. Morency, \"Multimodal machine learning: a survey and taxonomy\",   IEEE Trans. Pattern Anal. Mach. Intell. 41, 423\u2013443 (2019). doi:10.1109/TPAMI.2018.2798607</li> <li>W. Guo, J. Wang, S. Wang, \"Deep multimodal representation learning: a survey\", IEEE Access 7, 63373\u201363394   (2019). doi:10.1109/ACCESS.2019.2916887</li> <li>J. Ngiam et al., \"Multimodal deep learning\", ICML (2011).</li> <li>A. van den Oord, Y. Li, O. Vinyals, \"Representation learning with contrastive predictive coding\" (2018).   arXiv:1807.03748</li> <li>A. Radford et al., \"Learning transferable visual models from natural language supervision\" (CLIP), ICML   (2021). arXiv:2103.00020</li> <li>V. G. Satorras, E. Hoogeboom, M. Welling, \"E(n) equivariant graph neural networks\", ICML (2021).   arXiv:2102.09844</li> <li>V. Moro, C. Loh et al., \"Multimodal foundation models for material property prediction and discovery\",   Newton (2025). doi:10.1016/j.newton.2025.100016</li> <li>A. Babu, N. M. A. Krishnan, \"Multimodal and cross-modal learning techniques\", APL Mach. Learn. 4, 030901   (2026). doi:10.1063/5.0346744</li> <li>A. Babu, R. Almeida Gouv\u00eaa, G.-M. Rignanese, \"Toward automated discovery with generative models multimodal   learning and closed loop workflows in inverse materials design\", Cell Rep. Phys. Sci. 7, 103561 (2026).   doi:10.1016/j.xcrp.2026.103561</li> <li>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, \"MEIDNet: Multimodal generative AI framework for   inverse materials design\", npj Comput. Mater. (2026). doi:10.1038/s41524-026-02153-3</li> </ul>"},{"location":"learn/architectures.html","title":"Architecture Atlas","text":"<p>Five ways to build a model from several modalities, drawn the same way: inputs on the left, outputs on the right. Each card gives the idea, when it fits, its strengths and weaknesses, an example from materials science and references. The comparison table shows which of these designs MEIDNet uses, and the advisor suggests one for a given dataset and task.</p> <p>New to the topic? Start with Learn multimodality. Babu and Krishnan (2026) review multimodal and cross-modal learning techniques in more depth.</p> Modality A Modality B concatenateab one model prediction Modality A Modality B model A model B guess A guess B combine average or vote Structure Properties EGNN MLP shared space (128-d) align average decoders structure andproperties A tokens (atoms) B tokens (spectrum) querieskeys, values cross-attention fused tokens\u2192 prediction properties of the batch structures true pairs:pulled together other pairs:pushed apart"},{"location":"learn/architectures.html#early-fusion","title":"Early fusion","text":"<p>Idea. Join the modalities at the input: turn each into features, concatenate them into one vector, train one model on it.</p> <p>When it fits. Every material has every modality; the features have a fixed length and similar scales; you want a quick baseline.</p> <p>Strengths. Simple. The model can use interactions between features from the first layer. Weaknesses. A missing modality breaks the input. Very different shapes (a graph and a curve) are hard to concatenate. A large modality can drown a small one.</p> <p>Materials example. Composition descriptors, the processing temperature and a few peak positions from an XRD pattern, concatenated and given to a gradient-boosted regressor.</p> <p>References: Snoek et al. 2005; Baltru\u0161aitis et al. 2019.</p>"},{"location":"learn/architectures.html#late-fusion","title":"Late fusion","text":"<p>Idea. One model per modality, each makes its own prediction; the predictions are combined by averaging, voting or a small model on top.</p> <p>When it fits. Modalities are often missing, models for each already exist, or robustness matters more than squeezing out the last bit of accuracy.</p> <p>Strengths. Modular. Tolerates a missing modality. Each model can be trained on its own, larger dataset. Weaknesses. Cannot learn how the modalities interact. There is no shared representation, so no translation from one modality to another and no inverse design.</p> <p>Materials example. A structure-based graph network and a spectrum-based convolutional network each predict a phase label; the final label averages their probabilities.</p> <p>References: Snoek et al. 2005; Baltru\u0161aitis et al. 2019.</p>"},{"location":"learn/architectures.html#shared-latent","title":"Shared latent space (joint and coordinated)","text":"<p>Idea. One encoder per modality maps into a common space. In a coordinated representation the latents stay separate but are trained to agree; in a joint representation they are merged into one vector that the decoders read. MEIDNet does both: the structure latent and the property latent are aligned by contrastive learning, then averaged into the joint latent; the decoders read the joint latent and the property latent alone.</p> <p>When it fits. You want to translate between modalities (properties \u2192 structure for inverse design), retrieve one modality from another, or search a space that both understand. You have paired examples.</p> <p>Strengths. Translation in both directions, retrieval, and one space to optimise in. Each modality keeps the encoder that suits its shape. Weaknesses. Needs paired data. The space is only as good as the alignment. Several losses must be balanced.</p> <p>Materials example. MEIDNet: crystal structures and DFT properties of 18,928 cubic perovskites, searched for compositions with a target band gap and formation enthalpy (the paper's experiment).</p> <p>References: Ngiam et al. 2011; Baltru\u0161aitis et al. 2019; Babu et al. 2026 (MEIDNet); Moro, Loh et al. 2025; Babu and Krishnan 2026.</p>"},{"location":"learn/architectures.html#cross-attention","title":"Cross-attention and transformer fusion","text":"<p>Idea. Each modality becomes a sequence of tokens (atoms, segments of a spectrum, words). Attention lets every token of one modality look up the tokens of the other that matter to it, layer after layer.</p> <p>When it fits. The modalities have parts that interact in detail and are not aligned in advance; the dataset is large.</p> <p>Strengths. Learns which parts relate to which, and handles sequences of different lengths without manual alignment. Weaknesses. Needs much data and compute; the cost grows with the product of the two sequence lengths; harder to interpret than a single shared vector.</p> <p>Materials example. Spectrum segments attend to the atoms of the structure, so the model learns which sites shape which spectral features.</p> <p>References: Tsai et al. 2019 (Multimodal Transformer); Lu et al. 2019 (ViLBERT).</p>"},{"location":"learn/architectures.html#contrastive","title":"Contrastive learning","text":"<p>Idea. A training objective rather than a wiring. In a batch of paired examples, each example's two embeddings must be more similar to each other than to any other example's. It is how the latents of a shared space are coordinated.</p> <p>When it fits. You have pairs (a structure and its properties, a pattern and its structure) and want a space in which nearest neighbours are meaningful across modalities.</p> <p>Strengths. Learns from the pairing alone; provides cross-modal retrieval; scales to large datasets. Weaknesses. Needs large enough batches; sensitive to the temperature; two similar materials in the same batch are treated as negatives and pushed apart.</p> <p>In MEIDNet. Symmetric InfoNCE with temperature 0.01 and weight 5, switched on gradually over the warm-up (the paper's curriculum). The formula and a playground.</p> <p>References: van den Oord et al. 2018 (InfoNCE); Radford et al. 2021 (CLIP); Babu and Krishnan 2026.</p>"},{"location":"learn/architectures.html#side-by-side","title":"Side by side","text":"early fusion late fusion shared latent cross-attention contrastive (objective) where the modalities meet input features predictions a common latent space inside the network, token by token the latent space, through the loss every modality needed for every sample yes no paired examples for training usually paired examples learns interactions between modalities yes no yes yes, in detail for whole samples translates one modality into another no no yes with a decoder retrieval data needed small to medium small per model medium large medium to large used in MEIDNet \u2013 \u2013 yes: structure and property encoders, averaged latent \u2013 yes: symmetric InfoNCE <p>Beyond these five. Conditional generative models such as diffusion (MatterGen) or variational autoencoders (CDVAE) also translate from properties to structures, and they generate free atomic arrangements. MEIDNet instead decodes compositions onto a prototype family and checks them against chemical rules. Babu, Gouv\u00eaa and Rignanese (2026) review generative models, multimodal learning and closed-loop workflows for inverse materials design.</p>"},{"location":"learn/architectures.html#advisor","title":"Which architecture should I use?","text":"<p>Select the data you have and the task. The recommendation explains the choice and says where it can be run.</p> The advisor needs JavaScript. In short: structures with scalar properties, for design or retrieval \u2192 shared latent space with contrastive alignment (MEIDNet). Formula only \u2192 a composition model such as Roost. A pattern or spectrum to identify a phase \u2192 a 1D convolutional classifier. Modalities that are often missing \u2192 late fusion. Long sequences that interact in detail \u2192 cross-attention."},{"location":"learn/architectures.html#terminology","title":"A note on words","text":"<p>\"Early fusion\" in the MEIDNet paper names the model in which the structure latent and the property latent are averaged into one joint latent: <code>z_joint = (z_c + z_p) / 2</code> in <code>meidnet/model.py</code>. In the survey literature, early fusion usually means joining raw features at the input, as in the first card; averaging two learned latents, each from its own encoder, is intermediate or model-level fusion, the shared-latent card. Both names describe the same code. The checkpoints shipped here are the paper's early-fusion models, and the code implements only this fusion: there is no late-fusion option in MEIDNet.</p> <p>\"Curriculum\" in the paper is the contrastive warm-up: the alignment weight grows from zero to its full value over <code>training.contrastive_warmup_epochs</code>, so that the decoders learn first.</p>"},{"location":"learn/architectures.html#references","title":"References","text":"<ul> <li>C. G. M. Snoek, M. Worring, A. W. M. Smeulders, \"Early versus late fusion in semantic video analysis\",   ACM Multimedia (2005). doi:10.1145/1101149.1101236</li> <li>T. Baltru\u0161aitis, C. Ahuja, L.-P. Morency, \"Multimodal machine learning: a survey and taxonomy\",   IEEE TPAMI 41, 423\u2013443 (2019). doi:10.1109/TPAMI.2018.2798607</li> <li>J. Ngiam et al., \"Multimodal deep learning\", ICML (2011).</li> <li>Y.-H. H. Tsai et al., \"Multimodal Transformer for unaligned multimodal language sequences\", ACL (2019).   doi:10.18653/v1/P19-1656</li> <li>J. Lu, D. Batra, D. Parikh, S. Lee, \"ViLBERT: pretraining task-agnostic visiolinguistic representations for   vision-and-language tasks\", NeurIPS (2019). arXiv:1908.02265</li> <li>A. van den Oord, Y. Li, O. Vinyals, \"Representation learning with contrastive predictive coding\" (2018).   arXiv:1807.03748</li> <li>A. Radford et al., \"Learning transferable visual models from natural language supervision\", ICML (2021).   arXiv:2103.00020</li> <li>R. E. A. Goodall, A. A. Lee, \"Predicting materials properties without crystal structure: deep representation   learning from stoichiometry\" (Roost), Nat. Commun. 11, 6280 (2020). doi:10.1038/s41467-020-19964-7</li> <li>W. B. Park et al., \"Classification of crystal structure using a convolutional neural network\", IUCrJ 4,   486\u2013494 (2017). doi:10.1107/S205225251700714X</li> <li>C. Zeni et al., \"A generative model for inorganic materials design\" (MatterGen), Nature 639, 624\u2013632 (2025).   doi:10.1038/s41586-025-08628-5</li> <li>T. Xie et al., \"Crystal diffusion variational autoencoder for periodic material generation\" (CDVAE), ICLR   (2022). arXiv:2110.06197</li> <li>V. Moro, C. Loh et al., \"Multimodal foundation models for material property prediction and discovery\",   Newton (2025). doi:10.1016/j.newton.2025.100016</li> <li>A. Babu, N. M. A. Krishnan, \"Multimodal and cross-modal learning techniques\", APL Mach. Learn. 4, 030901   (2026). doi:10.1063/5.0346744</li> <li>A. Babu, R. Almeida Gouv\u00eaa, G.-M. Rignanese, \"Toward automated discovery with generative models multimodal   learning and closed loop workflows in inverse materials design\", Cell Rep. Phys. Sci. 7, 103561 (2026).   doi:10.1016/j.xcrp.2026.103561</li> <li>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, \"MEIDNet: Multimodal generative AI framework for   inverse materials design\", npj Comput. Mater. (2026). doi:10.1038/s41524-026-02153-3</li> </ul>"},{"location":"learn/recipes.html","title":"Recipes by problem","text":"<p>Start from a research question. Each recipe goes problem \u2192 architecture \u2192 why \u2192 data \u2192 implementation \u2192 tutorial \u2192 references. The implementation is MEIDNet where it applies; otherwise the recipe names established alternatives.</p> <p>For step-by-step changes to a MEIDNet configuration (another target, another family, a new rule), see the how-to guides.</p> problem architecture implementation Perovskites with a target band gap that are likely to be stable shared latent + contrastive MEIDNet Candidates from your own DFT results for one family shared latent + contrastive MEIDNet Screen a whole family for a target before searching structure encoder of the shared space MEIDNet Identify the crystal structure from an XRD pattern 1D convolutional network; contrastive pattern \u2194 structure 1D CNN classifiers; pattern encoder planned for MEIDNet Predict a property when only the formula is known composition network Roost and similar composition models Connect synthesis text with structures text encoder + contrastive or cross-attention text encoder planned for MEIDNet Design new atomic arrangements conditional generative model over full geometry MatterGen, CDVAE"},{"location":"learn/recipes.html#target-band-gap","title":"Perovskites with a target band gap that are likely to be stable","text":"Problem Find cubic ABX\u2083 halide perovskites with a direct band gap near 1.5 eV and a negative formation enthalpy. Architecture Shared latent space with contrastive alignment: MEIDNet. Why The question runs from properties to structure. A shared space whose crystal decoder is also trained from the property latent alone gives that direction, and the family's rules keep the chemistry sensible. Data Perov-5: 18,928 structures with DFT band gap and formation enthalpy (dataset card). The published checkpoint is trained on it. Implementation MEIDNet: <code>meidnet demo</code>, <code>meidnet generate</code> or the Studio's Targets and Search blocks. Tutorial The paper's experiment \u00b7 in the browser: Studio, Targets block \u00b7 screen the candidates with MACE References Babu et al. 2026 (MEIDNet); Castelli et al. 2012 (the dataset); Babu, Gouv\u00eaa and Rignanese 2026 (review of inverse design). <p>One command with the published model:</p> <pre><code>meidnet demo --family halide --band-gap 1.5 --enthalpy -0.10\n</code></pre> <p>Or in <code>meidnet.yaml</code> (from <code>examples/perov5/meidnet.yaml</code>):</p> <pre><code>generation:\n  family: perovskite_abx3\n  variant: halide\n  objectives:\n    - {property: dir_gap,  loss: l2, weight: 10000, select_weight: 1.0}\n    - {property: heat_all, loss: l1, weight: 6000,  select_weight: 0.4}\n  targets:\n    - {dir_gap: 1.5, heat_all: -0.10}\n</code></pre>"},{"location":"learn/recipes.html#own-family","title":"Candidates from your own DFT results for one family","text":"Problem You computed structures and properties for one prototype family, for example A\u2082BB\u2032X\u2086 double perovskites, and want new compositions with chosen property values. Architecture Shared latent space with contrastive alignment, trained on your table. Why The same reasoning as above. Every scalar column you give becomes a target you can set, and the family file says which sites exist and which elements may sit on them. Data A table with an id, numeric property columns and one CIF per row. Public sources by application are on Databases by application. Implementation MEIDNet: <code>meidnet init</code>, <code>check</code>, <code>train</code>, <code>generate</code>, or the Studio's Data and Model blocks. Tutorial Bring your own dataset \u00b7 in the browser: Studio, Data block \u00b7 change the material family References Babu et al. 2026 (MEIDNet). <pre><code>data:\n  table: my_results.csv\n  id_column: id\n  cif_column: cif\n  properties:\n    - {column: band_gap, unit: eV}\n    - {column: e_form,   unit: eV/atom}\ngeneration:\n  family: double_perovskite_a2bbx6\n  variant: halide\n  objectives:\n    - {property: band_gap, loss: l2,      weight: 10000}\n    - {property: e_form,   loss: at_most, weight: 6000}\n  targets:\n    - {band_gap: 1.8, e_form: -0.10}\n</code></pre> <p>Then <code>meidnet check</code> \u2192 <code>meidnet train</code> \u2192 <code>meidnet generate</code>; each step writes a report.</p>"},{"location":"learn/recipes.html#screen-family","title":"Screen a whole family for a target before searching","text":"Problem Before running a search, see every composition the family allows, its rule values and its predicted properties, and which are predicted closest to the target. Architecture The structure encoder of the shared space, read in the forward direction (structure \u2192 properties). Why For a prototype family the set of compositions is finite, so it can be scored exhaustively in seconds. Comparing these predictions with the search's candidates is a useful sanity check. Data A trained model and a family. Implementation MEIDNet: <code>meidnet space</code> or the Studio's design-space view. Tutorial Explore in 3D \u00b7 in the browser: Studio, design space \u00b7 How MEIDNet works References Babu et al. 2026 (MEIDNet). <pre><code>meidnet space examples/perov5/meidnet.yaml --model checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth -o space.csv\n</code></pre>"},{"location":"learn/recipes.html#xrd-structure","title":"Identify the crystal structure from an XRD pattern","text":"Problem Given a powder diffraction pattern, tell which crystal system, space group or known structure it comes from. Architecture A 1D convolutional classifier on the pattern (one modality). Multimodal: a pattern encoder aligned with a structure encoder by contrastive learning, so a pattern retrieves the structures it matches. Why A pattern is a fingerprint of the lattice. A classifier needs labels; a contrastive model needs only (pattern, structure) pairs, which can be simulated from any structure database. Data Simulate patterns from computed or experimental structures (semiconductor physics, all databases), for example with pymatgen's XRD calculator; validate on measured patterns. Implementation 1D convolutional classifiers such as Park et al. (2017). In MEIDNet, an encoder for binned XRD patterns is the first item of the roadmap and is not yet available. References Park et al. 2017; Baltru\u0161aitis et al. 2019; Babu and Krishnan 2026."},{"location":"learn/recipes.html#formula-only","title":"Predict a property when only the formula is known","text":"Problem Predict a band gap or a stability measure for compositions whose crystal structure is unknown, as in many experimental tables. Architecture A composition network that reads the formula as a weighted set of elements, such as Roost. With extra numeric inputs (temperature, doping), early fusion of composition features and those numbers is a strong baseline. Why MEIDNet needs a structure for every row. If your formulas all sit on one prototype (for example ABX\u2083), you can build those structures on the prototype and use MEIDNet; otherwise a composition model is the appropriate choice. Data Experimental band gaps and Matbench tasks are listed on Databases by application. Implementation Roost or another composition model; outside MEIDNet, which requires a crystal structure per material. References Goodall and Lee 2020 (Roost); Dunn et al. 2020 (Matbench)."},{"location":"learn/recipes.html#synthesis-text","title":"Connect synthesis text with structures","text":"Problem Link written synthesis procedures to the structures they produce, to suggest how a candidate might be made or which candidates resemble known syntheses. Architecture A text encoder (a language model) aligned with a structure encoder by contrastive learning; cross-attention when individual words must be tied to individual elements or steps. Why Text and structures have no common format; a shared space lets each be compared with the other. Data Text-mined synthesis recipes (Kononova et al. 2019) paired with structures from a structure database. Implementation A text encoder that maps into MEIDNet's shared space is planned and not yet available. References Kononova et al. 2019; Moro, Loh et al. 2025; Babu and Krishnan 2026."},{"location":"learn/recipes.html#new-arrangements","title":"Design new atomic arrangements","text":"Problem Generate crystals whose atomic arrangement is not one of a few known prototypes. Architecture A conditional generative model over the full geometry: diffusion (MatterGen) or a variational autoencoder (CDVAE). Why MEIDNet places compositions on a family prototype and does not invent arrangements (scope). Models that generate positions and cells directly can. Data Large structure databases; the generative benchmarks are listed under generative-model benchmarks. Implementation MatterGen, CDVAE and related generative models. References Zeni et al. 2025 (MatterGen); Xie et al. 2022 (CDVAE); Babu, Gouv\u00eaa and Rignanese 2026 (review)."},{"location":"learn/recipes.html#references","title":"References","text":"<ul> <li>A. Babu, R. Almeida Gouv\u00eaa, P. Vandergheynst, G.-M. Rignanese, \"MEIDNet: Multimodal generative AI framework for   inverse materials design\", npj Comput. Mater. (2026). doi:10.1038/s41524-026-02153-3</li> <li>I. E. Castelli et al., \"New cubic perovskites for one- and two-photon water splitting using the computational   materials repository\", Energy Environ. Sci. 5, 9034 (2012). doi:10.1039/C2EE22341D</li> <li>W. B. Park et al., \"Classification of crystal structure using a convolutional neural network\", IUCrJ 4,   486\u2013494 (2017). doi:10.1107/S205225251700714X</li> <li>R. E. A. Goodall, A. A. Lee, \"Predicting materials properties without crystal structure: deep representation   learning from stoichiometry\", Nat. Commun. 11, 6280 (2020). doi:10.1038/s41467-020-19964-7</li> <li>A. Dunn et al., \"Benchmarking materials property prediction methods: the Matbench   test set and Automatminer reference algorithm\", npj Comput. Mater. 6, 138 (2020). doi:10.1038/s41524-020-00406-3</li> <li>O. Kononova et al., \"Text-mined dataset of inorganic materials synthesis recipes\", Sci. Data 6, 203 (2019).   doi:10.1038/s41597-019-0224-1</li> <li>A. Babu, N. M. A. Krishnan, \"Multimodal and cross-modal learning techniques\", APL Mach. Learn. 4, 030901   (2026). doi:10.1063/5.0346744</li> <li>A. Babu, R. Almeida Gouv\u00eaa, G.-M. Rignanese, \"Toward automated discovery with generative models multimodal   learning and closed loop workflows in inverse materials design\", Cell Rep. Phys. Sci. 7, 103561 (2026).   doi:10.1016/j.xcrp.2026.103561</li> <li>C. Zeni et al., \"A generative model for inorganic materials design\", Nature 639, 624\u2013632 (2025).   doi:10.1038/s41586-025-08628-5</li> <li>T. Xie et al., \"Crystal diffusion variational autoencoder for periodic material generation\", ICLR (2022).   arXiv:2110.06197</li> <li>T. Baltru\u0161aitis, C. Ahuja, L.-P. Morency, \"Multimodal machine learning: a survey and taxonomy\", IEEE TPAMI 41, 423\u2013443 (2019). doi:10.1109/TPAMI.2018.2798607</li> <li>V. Moro, C. Loh et al., \"Multimodal foundation models for material property prediction and discovery\",   Newton (2025). doi:10.1016/j.newton.2025.100016</li> </ul>"},{"location":"recipes/add-constraint.html","title":"Add a rule (constraint)","text":"<p>A rule is a function that looks at a candidate and returns pass/fail, the measured value, the allowed window and a sentence. The search enforces it; every report and the Studio explain it.</p>"},{"location":"recipes/add-constraint.html#1-write-it","title":"1. Write it","text":"<p><code>my_rules.py</code>:</p> <pre><code>from meidnet.constraints import CONSTRAINTS\n\n@CONSTRAINTS.register(\"no_lead\", \"Rejects any composition containing Pb.\")\ndef no_lead(cand):\n    return cand.result(\"no_lead\", \"Pb\" not in cand.elements.values())\n\n@CONSTRAINTS.register(\"mass_window\", \"Mean atomic mass must lie between min and max (g/mol).\")\ndef mass_window(cand, min=0.0, max=200.0):\n    from pymatgen.core import Element\n    m = sum(Element(e).atomic_mass for e in cand.structure.species) / len(cand.structure)\n    return cand.result(\"mass_window\", min &lt;= m &lt;= max, float(m), (min, max), f\"mean mass {m:.1f}\")\n</code></pre> <p>What a candidate gives you: <code>cand.elements</code> (<code>{group: element}</code>), <code>cand.structure</code> (pymatgen, after symmetry refinement), <code>cand.raw</code> (before), <code>cand.lattice_a</code>, <code>cand.predictions</code> (<code>{property: value}</code>), <code>cand.radius(group)</code>.</p>"},{"location":"recipes/add-constraint.html#2-register-the-file-and-name-the-rule","title":"2. Register the file and name the rule","text":"<pre><code>plugins: [my_rules.py]\ngeneration:\n  family: my_perovskite.yaml      # a copy of the built-in family with the rules appended:\n</code></pre> <pre><code>constraints:\n  - {name: charge_neutrality}\n  - ...\n  - {name: no_lead}\n  - {name: mass_window, min: 0, max: 120}\n</code></pre>"},{"location":"recipes/add-constraint.html#3-run","title":"3. Run","text":"<p><code>meidnet generate meidnet.yaml</code>. The candidate cards list No lead and mass window with the measured value; the funnel shows how many compositions each removed.</p> <p>A complete example lives in <code>examples/custom_rules/</code> (a no toxic elements rule and a max cell edge rule), and notebook 03 does the same in Colab.</p>"},{"location":"recipes/add-constraint.html#soft-terms","title":"Soft terms","text":"<p>Rules decide acceptance. To steer the latent search instead (a penalty with a gradient), register a search term with <code>@SEARCH_TERMS.register</code> in <code>meidnet.terms</code>; it receives a <code>SearchContext</code> with the decoder's outputs. See rules and search terms.</p>"},{"location":"recipes/add-property.html","title":"Add a property","text":"<p>MEIDNet's property modality is a vector of any length. Adding a property is a data change, not a code change.</p> <ol> <li>Add the column to your table.</li> <li> <p>List it under <code>data.properties</code> (label and unit are for the reports; <code>normalize: true</code> is the default and    recommended when properties have different scales):</p> <pre><code>data:\n  properties:\n    - {column: band_gap,     label: Band gap,          unit: eV}\n    - {column: formation_e,  label: Formation energy,  unit: eV/atom}\n    - {column: dielectric,   label: Dielectric constant}\n</code></pre> </li> <li> <p><code>meidnet check</code> \u2192 <code>meidnet train</code>. The training report rates each property separately.</p> </li> <li>Use it in <code>generation.objectives</code> and <code>targets</code>.</li> </ol> <p>Properties with many missing values: rows with a missing value are skipped (the check report counts them). Fill them, or drop the property.</p> <p>The published model</p> <p>The Perov-5 checkpoint has exactly two properties (formation enthalpy, direct band gap) and cannot steer others; train your own model to add some.</p>"},{"location":"recipes/change-family.html","title":"Change the material family","text":"<p>A family file tells MEIDNet what kind of crystal to build. The built-in ones are <code>meidnet/families/perovskite_abx3.yaml</code> and <code>double_perovskite_a2bbx6.yaml</code>; <code>meidnet families &lt;name&gt;</code> prints a summary. To use your own, copy one, edit it, and point to it:</p> <pre><code>generation:\n  family: my_family.yaml\n  variant: null          # or a variant you define\n</code></pre>"},{"location":"recipes/change-family.html#the-parts-of-a-family-file","title":"The parts of a family file","text":"<pre><code>name: spinel_ab2x4\ntitle: Normal spinel AB2X4 (Fd-3m, primitive cell)\n\nprototype:                    # WHERE atoms sit (fractional coordinates, in decoder slot order)\n  lattice: fcc_primitive      # cubic | fcc_primitive | bcc_primitive | explicit a,b,c,alpha,beta,gamma\n  reference_a: 8.1\n  sites:\n    - {group: A, frac: [0.125, 0.125, 0.125]}\n    - ...\n\ngroups:                       # WHO may sit there, and with which charges\n  A: {description: tetrahedral cation, elements: [Mg, Zn, Fe, Co, Ni, Mn], oxidation_states: {Mg: [2], ...}}\n  B: {...}\n  X: {...}\n\nsampling_order: [B, X, A]     # the LAST group is restricted to charge-balancing elements\n\nlattice:                      # HOW BIG the cell is\n  rule: bond_sum              # a = factor \u00d7 \u03a3 ionic radii of the listed groups (or rule: fixed)\n  groups: [B, X]\n  factor: 2.0\n  min: 3.0\n  max: 12.0\n\nvariants:                     # named sub-families: element subsets and rule parameters\n  oxide: {groups: {X: {elements: [O]}}, params: {tolerance_factor: {min: 0.8, max: 1.05}}}\n\nsearch_terms:                 # soft penalties during the latent search (see reference/rules)\n  - {name: site_separation, weight: 200}\n  - ...\n\nconstraints:                  # hard rules every saved candidate must pass\n  - {name: charge_neutrality}\n  - {name: min_distance, min: 0.8}\n  - ...\n</code></pre>"},{"location":"recipes/change-family.html#alignment-of-training-data","title":"Alignment of training data","text":"<p>For generation to mean anything, the model must have been trained on structures whose atoms are in the prototype's site order. <code>data.align_to_prototype: true</code> (default) re-orders every training structure to match the family named in <code>generation.family</code> (or top-level <code>family:</code>); structures that do not fit are reported by <code>meidnet check</code>. Raise <code>data.prototype_tolerance</code> for distorted cells.</p>"},{"location":"recipes/change-family.html#checking-a-family-file","title":"Checking a family file","text":"<pre><code>meidnet families my_family.yaml --variant oxide\nmeidnet space meidnet.yaml      # every composition with rule values \u2192 CSV\n</code></pre> <p>Errors name the problem: unknown element symbols, a prototype group without a definition, an override that matches no rule, a group left empty after your filters.</p>"},{"location":"recipes/change-family.html#worked-example","title":"Worked example","text":"<p>The double perovskite page walks through a 10-site, non-cubic-primitive family written this way.</p>"},{"location":"recipes/change-targets.html","title":"Change the target properties","text":""},{"location":"recipes/change-targets.html#a-different-value","title":"A different value","text":"<pre><code>generation:\n  targets:\n    - {band_gap: 1.2}\n    - {band_gap: 1.8}      # each entry is its own search\n</code></pre>"},{"location":"recipes/change-targets.html#a-different-kind-of-goal","title":"A different kind of goal","text":"<p><code>loss</code> says how a prediction is compared with the target:</p> <code>loss</code> meaning typical use <code>l2</code> squared distance (default) hit a value <code>l1</code> absolute distance hit a value, less sensitive to outliers <code>at_most</code> only values above the target are penalised formation enthalpy \u2264 \u22120.1 eV/atom <code>at_least</code> only values below are penalised dielectric constant \u2265 20 <pre><code>objectives:\n  - {property: band_gap,   loss: l2,       weight: 10000, select_weight: 1.0}\n  - {property: enthalpy,   loss: at_most,  weight: 6000,  select_weight: 0.4}\n</code></pre> <p><code>weight</code> acts during the latent search; <code>select_weight</code> when ranking finished candidates.</p>"},{"location":"recipes/change-targets.html#a-different-property","title":"A different property","text":"<p>The model knows the properties it was trained on (<code>meidnet info model.pt</code> lists them). To steer a property that is not in the model, add it to the data and retrain.</p>"},{"location":"recipes/change-targets.html#in-the-studio","title":"In the Studio","text":"<p>Drag the target slider; the predicted-closest table and the scatter update at once, and a warning appears if the target lies outside the training range. Export the YAML when you are happy.</p>"},{"location":"recipes/elements.html","title":"Exclude or restrict elements","text":"<pre><code>generation:\n  exclude_elements: [Pb, Cd, Hg]           # never used, on any site\n  only_elements:                           # restrict a site group to a shortlist\n    B: [Ti, Zr, Hf, Sn]\n</code></pre> <p>Both are applied to the family before the search starts, so the decoder cannot even propose an excluded element. The Studio's element chips do the same thing (click to toggle) and show the counts change.</p> <p>To change which elements a family allows at all, or their oxidation states, copy the family file and edit the <code>groups:</code> section \u2014 see Change the material family.</p>"},{"location":"recipes/screen.html","title":"Screen stability with MACE","text":"<pre><code>pip install \"meidnet[stability]\"            # ase + mace-torch (~500 MB of weights on first use)\nmeidnet screen runs/my_oxides/generation --train-csv materials.csv\n</code></pre> <p>For each CIF: relax cell and positions with the MACE-MP-0 universal potential, compute the formation energy per atom against elemental reference phases, and report</p> <ul> <li>S stable \u2014 \u0394H_f \u2264 0.10 eV/atom (Materials Project's \"potentially synthesisable\" criterion; <code>--threshold</code>),</li> <li>U unique \u2014 no structural duplicate among the candidates,</li> <li>N novel \u2014 reduced formula absent from the training table.</li> </ul> <p>Results go to <code>stability.csv</code> in the folder. Structures with |\u0394H_f| &gt; 5 eV/atom are flagged as artefacts and excluded from the rates. The reference phases and GGA-style O/N corrections are those of the paper.</p> <p>This is a first filter, not a verdict: MACE formation energies carry ~0.1 eV/atom uncertainty, and a cubic prototype may relax to a lower-symmetry structure.</p>"},{"location":"reference/api.html","title":"Python API","text":"<p>The command line is a thin layer over these functions; use them from notebooks or your own scripts.</p> <pre><code>from meidnet.config import load_config, config_from_dict\nfrom meidnet.pipeline import check, train, generate\n\ncfg = load_config(\"meidnet.yaml\")\ninfo = check(cfg)                 # dict: report (counts, skip reasons), records, family, report_path\ninfo = check(cfg, keep_structures=True)   # records also keep their pymatgen Structure (record.structure)\nmodel_path = train(cfg)           # writes model.pt + training_report.html, returns the path\nresult = generate(cfg)            # GenerationResult: .saved (candidates), .targets (funnel logs)\n</code></pre>"},{"location":"reference/api.html#pieces","title":"Pieces","text":"<pre><code>from meidnet.checkpoint import load_checkpoint, describe, property_ranges\nlm = load_checkpoint(\"model.pt\")          # LoadedModel: .model, .stats (property names/units/normalisation)\nlm.model.encode_crystal(x)                # structure \u2192 latent (x from meidnet.data.featurize)\nlm.model.encode_properties(p)             # normalised property vector \u2192 latent\n\nfrom meidnet.family import load_family\nfam = load_family(\"perovskite_abx3\", variant=\"halide\", exclude=[\"Pb\"], overrides={\"tolerance_factor\": {\"max\": 1.0}})\n\nfrom meidnet.constraints import build_candidate, evaluate, CONSTRAINTS\ncand = evaluate(build_candidate(fam, {\"A\": \"Cs\", \"B\": \"Sn\", \"X\": \"Br\"}), fam.constraints)\ncand.passed, [r.to_dict() for r in cand.results]\n\nfrom meidnet.generate import Designer\nfrom meidnet.config import GenerationSection\ng = GenerationSection(family=\"perovskite_abx3\", variant=\"halide\",\n                      objectives=[{\"property\": \"dir_gap\"}], targets=[{\"dir_gap\": 2.0}], rounds=3, steps=200)\nres = Designer(lm, fam, g, on_saved=lambda c, t: print(c.formula)).run(\"out/\")\n\nfrom meidnet.designspace import enumerate_space, space_to_csv\nspace = enumerate_space(fam, lm)          # every composition: rule values + predictions\n</code></pre> <p>Custom rules and search terms: <code>CONSTRAINTS.register</code>, <code>meidnet.terms.SEARCH_TERMS.register</code>.</p>"},{"location":"reference/api.html#extra-rules-in-the-configuration","title":"Extra rules in the configuration","text":"<p><code>generation.extra_constraints</code> adds rules to the family's own for one run, without editing the family file \u2014 this is how the Studio saves a predicted-property window:</p> <pre><code>generation:\n  family: perovskite_abx3\n  variant: halide\n  extra_constraints:\n    - {name: property_window, property: dir_gap, min: 1.5, max: 3.0}\n</code></pre> <p><code>meidnet.pipeline.family_for(cfg, need_variant=True)</code> returns the family a run uses, with these rules appended; <code>generation.family</code> / <code>generation.variant</code> take precedence over the top-level <code>family</code>.</p>"},{"location":"reference/api.html#3d-datasets-chemiscope","title":"3D datasets (chemiscope)","text":"<p><code>meidnet.studio.chemiscope</code> turns MEIDNet objects into chemiscope datasets \u2014 plain dicts you can save with <code>json.dump</code> and open at chemiscope.org, or show in a notebook with <code>chemiscope.show(...)</code>:</p> <pre><code>from meidnet.studio.chemiscope import (space_dataset, candidates_dataset, records_dataset,\n                                       published_data_dataset, structure_to_chemiscope)\n\nds = space_dataset(fam, space, lm, max_structures=2000)   # design space: formula, site_*, pred_*, rule_*, passes_all\nds = candidates_dataset(candidate_dicts, fam, run_dir, lm=lm)   # candidates of a run (reads their CIFs when present;\n                                                                # lm adds units and property names)\nds = records_dataset(info[\"records\"], cfg.data.properties, source=\"my data\")   # needs check(cfg, keep_structures=True)\nstructure_to_chemiscope(structure)                         # one pymatgen Structure \u2192 {size, names, x, y, z, cell}\n</code></pre> <p><code>candidate_dicts</code> are saved candidates as dicts \u2014 <code>[c.to_dict() for c in res.saved]</code> after a <code>Designer</code> run, or <code>[c for t in json.load(open(\"generation.json\"))[\"targets\"] for c in t[\"saved\"]]</code> from a finished run. <code>run_dir</code> is the run folder that contains <code>generation/</code> (<code>cfg.out_dir</code> after <code>meidnet generate</code>); where a CIF file is not found, the structure is rebuilt on the prototype from the candidate's elements.</p>"},{"location":"reference/api.html#scripting-the-studio","title":"Scripting the Studio","text":"<p>The Studio server is a small class you can drive without a browser \u2014 the 04 Colab notebook does exactly this. Every method takes the same JSON-like dicts the page sends; <code>session</code> keeps users (or experiments) apart.</p> <pre><code>import base64\nfrom meidnet.studio.server import Studio\n\nst = Studio(None)                                   # the published model; or Studio(load_config(\"meidnet.yaml\"))\nsid = \"my-session\"\nst.state(sid)                                       # model, ranges, families, limits, session info\n\nb64 = base64.b64encode(open(\"materials.csv\", \"rb\").read()).decode()\nup = st.upload({\"session\": sid, \"files\": [{\"name\": \"materials.csv\", \"b64\": b64}]})\nup[\"suggest\"]                                       # suggested id / CIF / property columns\n\nst.check_data({\"session\": sid, \"id_column\": \"material_id\", \"cif_column\": \"cif\",\n               \"properties\": [{\"column\": \"dir_gap\", \"unit\": \"eV\"}],\n               \"family\": \"perovskite_abx3\", \"variant\": \"halide\"})\nst.start_train({\"session\": sid, \"epochs\": 10, \"batch_size\": 16})\nst.status(sid, \"train\")                             # poll: running, progress (epoch, loss, curve), done, error\nst.select_model(sid, \"published\")                   # or \"own\"\n\nst.family({\"session\": sid, \"family\": \"perovskite_abx3\", \"variant\": \"halide\"})   # design space, scored\nst.start_search({\"session\": sid, \"generation\": {\"per_target\": 2, \"rounds\": 2, \"steps\": 100}})\nst.status(sid, \"search\")                            # candidates as they are found\nst.chemiscope(sid, \"space\", \"perovskite_abx3\", \"halide\")   # or \"candidates\" / \"data\"\nst.validate({\"yaml\": \"generation:\\n  rounds: 2\\n\"})        # {ok, generation, notes} or {ok: False, errors}\nst.export({\"session\": sid, \"generation\": {}})              # the meidnet.yaml text\n</code></pre>"},{"location":"reference/cli.html","title":"Command line","text":"<pre><code>meidnet init            write a starter meidnet.yaml next to your data\nmeidnet check  CONFIG   is my data usable?            \u2192 check_report.html\nmeidnet train  CONFIG   learn the latent space        \u2192 model.pt, training_report.html\nmeidnet generate CONFIG design candidates             \u2192 CIFs, generation_report.html\nmeidnet studio [CONFIG] interactive workbench in your browser\nmeidnet space  CONFIG   every composition of the family with rule values and predictions \u2192 CSV\nmeidnet demo            quick demo with the published perovskite model\nmeidnet screen DIR      stability screening of CIFs with MACE (optional extra)\nmeidnet info MODEL      describe a checkpoint\nmeidnet families [NAME] list / describe material families\nmeidnet schema          JSON Schema of the config (for editors and the Studio)\nmeidnet download-data   fetch the Perov-5 dataset used in the paper\n</code></pre> <p><code>meidnet &lt;command&gt; --help</code> lists the options. Useful ones:</p> <code>init --template perov5</code> the paper's configuration instead of a blank one <code>train --epochs N</code> override <code>training.epochs</code> <code>generate --quick</code> few rounds and steps \u2014 a smoke test <code>generate --model PATH</code> use another checkpoint <code>studio --export-static FILE.html</code> self-contained Studio for a website <code>studio --port 8765 --no-open</code> choose the port; do not open a browser <code>studio --public --host 0.0.0.0</code> host the Studio for other people: private session per visitor, capped budgets <code>studio --run-root DIR --docs-dir site</code> where sessions and searches are written; serve a built docs site at <code>/docs/</code> <code>space --chemiscope FILE.json</code> also write the design space as a 3D dataset for chemiscope.org <code>demo --family oxide --band-gap 3 --enthalpy -0.2 -n 5</code> demo targets <p>If <code>meidnet</code> is not on your PATH, <code>python -m meidnet.cli \u2026</code> is equivalent.</p>"},{"location":"reference/config.html","title":"Configuration reference","text":"<p>Every setting of <code>meidnet.yaml</code>, generated from the code (<code>meidnet schema</code> gives the JSON Schema). Unknown keys are rejected with a message naming the key, so typos cannot silently change a run.</p>"},{"location":"reference/config.html#top-level","title":"Top level","text":"setting type default meaning <code>name</code> string <code>meidnet_run</code> Name of this project; outputs go to output_dir. <code>description</code> string Free text shown at the top of reports. <code>output_dir</code> string <code>runs/{name}</code> Where checkpoints, CIFs and reports are written. <code>family</code> string/null <code>None</code> Family used to align training structures (defaults to generation.family). <code>model_path</code> string/null <code>None</code> Existing checkpoint to generate from (skip training). Default: /model.pt <code>plugins</code> array Python files that register your own constraints or search terms (see 'Add a constraint'). <code>data</code> object/null <code>None</code> <code>model</code> section <code>training</code> section <code>generation</code> object/null <code>None</code>"},{"location":"reference/config.html#data","title":"<code>data:</code>","text":"setting type default meaning <code>table</code> string required Path to a CSV/Excel/JSON table with one row per material. <code>id_column</code> string <code>material_id</code> Column with a unique identifier for each material. <code>cif_column</code> string/null <code>cif</code> Column that contains the CIF text of each structure. Set to null if you use structures_dir. <code>structures_dir</code> string/null <code>None</code> Folder with one CIF file per material, named .cif (used when there is no cif_column). <code>properties</code> array required Numeric columns that form the property modality (any number, at least one). <code>val_table</code> string/null <code>None</code> Optional separate validation table with the same columns. <code>val_fraction</code> number <code>0.1</code> Fraction of rows held out for validation when no val_table is given (0 = no validation). <code>max_sites</code> integer <code>20</code> Largest number of atoms per cell. Bigger structures are skipped. <code>neighbor_cutoff</code> number <code>4.0</code> Distance (\u00c5) below which two atoms count as neighbours. <code>align_to_prototype</code> boolean <code>True</code> Re-order the atoms of every structure so they match the family prototype's site order (needed for meaningful generation). The published Perov-5 model was trained without it. <code>prototype_tolerance</code> number <code>0.15</code> How far (fraction of a cell edge) an atom may sit from its prototype site and still count as that site. Raise it to accept distorted structures; lower it to keep only ideal ones."},{"location":"reference/config.html#dataproperties","title":"<code>data.properties[]</code>","text":"setting type default meaning <code>column</code> string required Name of the column in your table that holds this property. <code>label</code> string/null <code>None</code> Human-readable name used in reports (defaults to the column name). <code>unit</code> string Unit shown in reports, e.g. 'eV' or 'eV/atom'. <code>normalize</code> boolean <code>True</code> Standardise the property to zero mean and unit spread before training. Keep this on when properties have very different scales."},{"location":"reference/config.html#model","title":"<code>model:</code>","text":"setting type default meaning <code>latent_dim</code> integer <code>128</code> Size of the shared latent space where structures and properties meet. <code>node_hidden_dim</code> integer <code>128</code> Width of the per-atom features inside the graph network. <code>edge_dim</code> integer <code>64</code> Width of the per-bond messages inside the graph network. <code>species_embedding_dim</code> integer <code>64</code> Size of the learned element embedding. <code>property_hidden_dim</code> integer <code>128</code> Width of the property encoder's hidden layer. <code>decoder_coordinate_input</code> <code>data</code> / <code>zeros</code> / <code>prototype</code> <code>data</code> What the crystal decoder receives as starting atom positions during training. 'data' (published model) uses the true positions; generation always starts from 'zeros' or the prototype. 'zeros'/'prototype' make training and generation consistent (experimental)."},{"location":"reference/config.html#training","title":"<code>training:</code>","text":"setting type default meaning <code>epochs</code> integer <code>200</code> Number of passes over the training data. <code>batch_size</code> integer <code>16</code> Materials per training step (use 8 if memory is short). <code>learning_rate</code> number <code>0.001</code> Adam learning rate. <code>contrastive_weight</code> number <code>5.0</code> Strength of the alignment between structure and property latents. <code>temperature</code> number <code>0.01</code> Sharpness of the contrastive (InfoNCE) alignment. <code>contrastive_warmup_epochs</code> integer/null <code>None</code> Epochs over which the alignment strength ramps up from 0. Default: 60% of the epochs (the published model used 1200 of 2000). <code>loss_weights</code> section <code>seed</code> integer <code>0</code> Random seed for weight initialisation and data shuffling. <code>device</code> string <code>auto</code> 'auto', 'cpu' or 'cuda'. <code>save_every</code> integer <code>50</code> Write an intermediate checkpoint every N epochs."},{"location":"reference/config.html#trainingloss_weights","title":"<code>training.loss_weights:</code>","text":"setting type default meaning <code>joint_reconstruction</code> number <code>1.0</code> Rebuild the crystal from the joint (structure+property) latent. <code>joint_property</code> number <code>1.0</code> Predict the properties from the joint latent. <code>property_reconstruction</code> number <code>0.5</code> Rebuild the crystal from the property latent alone - this is what makes inverse design possible. <code>property_property</code> number <code>0.5</code> Predict the properties back from the property latent."},{"location":"reference/config.html#generation","title":"<code>generation:</code>","text":"setting type default meaning <code>family</code> string <code>perovskite_abx3</code> Built-in family name or path to your own family .yaml file. <code>variant</code> string/null <code>None</code> Sub-family, e.g. 'halide' or 'oxide' for perovskites. <code>objectives</code> array required Which properties to steer and how. <code>targets</code> array required One entry per design target, e.g. {dir_gap: 1.5, heat_all: -0.1}. Each runs its own search. <code>per_target</code> integer <code>4</code> Candidates to save for each target. <code>population</code> integer <code>48</code> Latent vectors optimised in parallel in each round. <code>rounds</code> integer <code>20</code> Maximum search rounds per target. <code>steps</code> integer <code>800</code> Gradient steps per round. <code>learning_rate</code> number <code>0.0012</code> Step size of the latent search. <code>anchor_weight</code> number <code>12.0</code> Keeps the search close to the latent that the target properties map to. <code>temperature_start</code> number <code>1.6</code> Softness of element choices at the start of a round. <code>temperature_end</code> number <code>0.9</code> Softness of element choices at the end of a round. <code>diversity_weight</code> number <code>0.9</code> Pushes the parallel searches apart from each other. <code>diversity_tau</code> number <code>0.25</code> <code>history_weight</code> number <code>1.0</code> Pushes new searches away from latents that already produced a saved candidate. <code>history_tau</code> number <code>0.25</code> <code>restart_patience</code> integer <code>250</code> Steps without improvement before a search is restarted with noise. <code>restart_noise</code> number <code>0.35</code> <code>grad_clip</code> number <code>1.0</code> <code>z_clip</code> number <code>5.0</code> <code>init_sigma</code> number <code>0.35</code> Noise added to the starting latents. <code>init_anchor_mix</code> number <code>0.75</code> Share of the property-encoder latent in the starting point. <code>init_rff_mix</code> number <code>0.6</code> Share of the target-dependent random-feature direction in the starting point. <code>init_noise_mix</code> number <code>0.25</code> Share of random noise in the starting point. <code>rff_frequencies</code> integer <code>48</code> <code>decode_temperature</code> number <code>1.25</code> Randomness when turning element scores into a choice. <code>decode_topk</code> integer <code>12</code> Only the k best-scoring elements of a group can be chosen. <code>decode_tries</code> integer <code>12</code> Attempts per latent to find a composition that passes all constraints. <code>anti_repeat_alpha</code> number <code>0.6</code> Down-weights elements that were already used in saved candidates. <code>anti_repeat_group</code> string/null <code>None</code> Group the anti-repeat applies to (default: first sampled group). <code>geometry_scale</code> number <code>1.0</code> Global strength of the family's search terms. <code>property_first</code> boolean <code>False</code> Ignore the family's search terms during the search (properties only). <code>dedup_formula</code> boolean <code>True</code> Save each composition at most once. <code>min_cosine_sep</code> number <code>0.985</code> Skip candidates whose latent is this similar (cosine) to an already saved one. <code>unique_decimals</code> integer <code>3</code> <code>exclude_elements</code> array Elements never to use, e.g. [Pb, Cd]. <code>only_elements</code> object Restrict groups to these elements, e.g. {B: [Ti, Zr, Hf]}. <code>overrides</code> object Change parameters of the family's search terms or constraints, e.g. {tolerance_factor: {max: 1.0}}. <code>extra_constraints</code> array Additional rules appended to the family's constraints, e.g. [{name: property_window, property: dir_gap, min: 1.0, max: 3.0}]. <code>disabled_rules</code> array Rules to switch off for this run, by name (or id when two rules share a name), e.g. [charge_neutrality]. The generation report lists them. <code>seed</code> integer <code>937</code> Random seed of the search. <code>amp</code> boolean <code>True</code> Use mixed precision on GPUs. <code>output_prefix</code> string/null <code>None</code> File-name prefix of saved CIFs (default: family variant)."},{"location":"reference/config.html#generationobjectives","title":"<code>generation.objectives[]</code>","text":"setting type default meaning <code>property</code> string required Property (column name) this objective refers to. <code>loss</code> <code>l2</code> / <code>l1</code> / <code>at_most</code> / <code>at_least</code> <code>l2</code> How a prediction is compared with the target: l2 = squared distance, l1 = absolute distance, at_most = only values above the target are penalised, at_least = only values below. <code>weight</code> number <code>10000.0</code> Importance of this objective during the latent search. <code>select_weight</code> number <code>1.0</code> Importance of this objective when ranking finished candidates. <code>select_loss</code> string/null <code>None</code> Comparison used for ranking (defaults to 'loss')."},{"location":"reference/families.html","title":"Family files","text":"<p>A family file is YAML with these keys. <code>meidnet families NAME</code> prints a summary and the constraints with their explanations; errors name the offending key or element.</p> key required meaning <code>name</code>, <code>title</code>, <code>description</code> name identification; <code>description</code> appears in the Studio <code>prototype.lattice</code> yes <code>cubic</code>, <code>fcc_primitive</code>, <code>bcc_primitive</code>, or explicit <code>a, b, c, alpha, beta, gamma</code> <code>prototype.reference_a</code> typical cubic cell edge (\u00c5), used by the <code>cubic_cell</code> search term <code>prototype.sites[]</code> yes <code>{group, frac}</code> in decoder slot order; fractional coordinates <code>groups.&lt;G&gt;.description</code> sentence shown in reports <code>groups.&lt;G&gt;.elements</code> yes every element the group may hold (all variants) <code>groups.&lt;G&gt;.sample</code> subset offered to the final element choice (default: <code>elements</code>) <code>groups.&lt;G&gt;.oxidation_states</code> <code>{element: [charges]}</code> used for charge balance and neutrality bias <code>sampling_order</code> order in which groups are chosen; the last is restricted to charge-balancing elements <code>lattice.rule</code> <code>bond_sum</code> (<code>a = factor \u00d7 \u03a3 r_ionic</code> of <code>groups</code>, clipped to <code>min..max</code>) or <code>fixed</code> <code>variants.&lt;V&gt;.groups</code> per-variant <code>elements</code>, <code>sample</code>, <code>oxidation_states</code> of any group <code>variants.&lt;V&gt;.params</code> <code>{rule_or_term_name: {param: value}}</code> applied to search terms and constraints <code>require_variant</code> default true: a variant must be chosen when variants exist <code>search_terms[]</code> <code>{name, \u2026params}</code> soft penalties during the latent search <code>logit_transforms[]</code> <code>{name, \u2026params}</code> edits of the decoder's element scores (e.g. <code>neutrality_bias</code>) <code>constraints[]</code> <code>{name, \u2026params}</code> hard rules; <code>min_distance</code> runs before symmetry refinement, the rest after <code>refine_symmetry</code>, <code>symprec</code> run spglib refinement on candidates (default true, 0.05) <p>User-side changes without editing the file: <code>generation.variant</code>, <code>exclude_elements</code>, <code>only_elements</code>, <code>overrides</code> (<code>{name: {param: value}}</code> for any search term or constraint).</p> <p>The built-in files are the complete reference: <code>perovskite_abx3.yaml</code>, <code>double_perovskite_a2bbx6.yaml</code>.</p>"},{"location":"reference/rules.html","title":"Rules and search terms","text":"<p>Rules (hard constraints) decide whether a candidate is saved. Search terms (soft penalties) steer the latent search towards decodable, sensible outputs. Both are referenced by name in family files.</p>"},{"location":"reference/rules.html#built-in-rules","title":"Built-in rules","text":""},{"location":"reference/rules.html#bond_window-sensible-fromto-bonds","title":"<code>bond_window</code> \u2014 Sensible {from}\u2013{to} bonds","text":"<p>Every {from} atom needs a {to} neighbour at {low}\u2013{high} \u00d7 the sum of their ionic radii: bonds may be neither squashed nor stretched.</p>"},{"location":"reference/rules.html#charge_neutrality-charge-balance","title":"<code>charge_neutrality</code> \u2014 Charge balance","text":"<p>The formal charges must add up to zero for at least one combination of the listed oxidation states (e.g. Cs\u207a Pb\u00b2\u207a I\u207b\u00d73 = 0).</p>"},{"location":"reference/rules.html#min_distance-no-overlapping-atoms","title":"<code>min_distance</code> \u2014 No overlapping atoms","text":"<p>Two atoms closer than {min} \u00c5 would overlap, which is physically impossible.</p>"},{"location":"reference/rules.html#octahedral_factor-octahedral-factor","title":"<code>octahedral_factor</code> \u2014 Octahedral factor","text":"<p>\u03bc = r_B / r_X says whether the B cation is large enough to hold six X anions around it ({min} \u2264 \u03bc \u2264 {max}).</p>"},{"location":"reference/rules.html#property_window-predicted-property-window","title":"<code>property_window</code> \u2014 Predicted {property} window","text":"<p>The model's predicted {property} must lie between {min} and {max}.</p>"},{"location":"reference/rules.html#tolerance_factor-goldschmidt-tolerance-factor","title":"<code>tolerance_factor</code> \u2014 Goldschmidt tolerance factor","text":"<p>t = (r_A + r_X) / (\u221a2 (r_B + r_X)) measures how well the A cation fits the cage of B\u2013X octahedra. Cubic perovskites form roughly for {min} \u2264 t \u2264 {max}.</p>"},{"location":"reference/rules.html#search-terms","title":"Search terms","text":"<ul> <li><code>cubic_cell</code> \u2014 Pulls the decoded cell towards a cube of the family's reference volume.</li> <li><code>entropy_bonus</code> \u2014 Rewards uncertainty early in a round so the search explores (fades to zero).</li> <li><code>group_consistency</code> \u2014 Sites of the same group should agree on their element (ramps up during a round).</li> <li><code>prototype_alignment</code> \u2014 Pulls decoded positions onto the prototype's site positions.</li> <li><code>short_distance_penalty</code> \u2014 Large penalty if any two decoded atoms come closer than <code>min</code>.</li> <li><code>site_repulsion</code> \u2014 Short-range exponential repulsion between decoded atoms.</li> <li><code>site_separation</code> \u2014 Keeps decoded atoms from collapsing onto each other (1/distance).</li> <li><code>soft_charge</code> \u2014 Expected total charge of the most likely composition should be zero (ramps up during a round).</li> <li><code>tolerance_penalty</code> \u2014 Soft version of the Goldschmidt tolerance-factor window, using the most likely elements.</li> </ul>"},{"location":"reference/rules.html#logit-transforms","title":"Logit transforms","text":"<ul> <li><code>neutrality_bias</code> \u2014 Raises the scores of anions whose charge can neutralise the most likely cations.</li> </ul>"},{"location":"reference/rules.html#your-own-rule","title":"Your own rule","text":"<pre><code>from meidnet.constraints import CONSTRAINTS\n\n@CONSTRAINTS.register(\"no_lead\", \"Rejects any composition containing Pb.\")\ndef no_lead(cand):\n    return cand.result(\"no_lead\", \"Pb\" not in cand.elements.values())\n</code></pre> <p>Save it as <code>my_rules.py</code>, list it under <code>plugins:</code> in <code>meidnet.yaml</code>, and add <code>- {name: no_lead}</code> to the family's <code>constraints</code>. See Add a rule.</p>"},{"location":"research/paper.html","title":"Paper","text":"<p>MEIDNet: Multimodal generative AI framework for inverse materials design. Anand Babu, Rog\u00e9rio Almeida Gouv\u00eaa, Pierre Vandergheynst, Gian-Marco Rignanese. npj Computational Materials (2026). doi:10.1038/s41524-026-02153-3 \u00b7 arXiv:2601.22009</p> <pre><code>@article{meidnet2026,\n  title   = {MEIDNet: Multimodal generative AI framework for inverse materials design},\n  author  = {Anand Babu and Rog{\\'e}rio Almeida Gouv{\\^e}a and Pierre Vandergheynst and Gian-Marco Rignanese},\n  journal = {npj Computational Materials},\n  year    = {2026},\n  doi     = {10.1038/s41524-026-02153-3}\n}\n</code></pre>"},{"location":"research/paper.html#code-and-data-of-the-paper","title":"Code and data of the paper","text":"<ul> <li>Code as published: tag <code>v1.0.0-paper</code>   (frozen; MEIDNet 2 reproduces it bit for bit \u2014 see Perov-5).</li> <li>Checkpoints: <code>checkpoints/</code> in the repository (2.8 MB each).</li> <li>Data: Perov-5 in the CDVAE split (<code>meidnet download-data</code>); Castelli et al., Energy Environ. Sci. 5, 5814 (2012);   Xie et al., ICLR (2022).</li> <li>Generated structures: <code>examples/perov5/paper_results/</code>.</li> </ul>"},{"location":"research/paper.html#related","title":"Related","text":"<p>Reviews by the authors:</p> <ul> <li>A. Babu, N. M. A. Krishnan, Multimodal and cross-modal learning techniques, APL Machine Learning 4, 030901   (2026). doi:10.1063/5.0346744</li> <li>A. Babu, R. Almeida Gouv\u00eaa, G.-M. Rignanese, Toward automated discovery with generative models multimodal   learning and closed loop workflows in inverse materials design, Cell Reports Physical Science 7, 103561   (2026). doi:10.1016/j.xcrp.2026.103561</li> </ul> <p>Methods MEIDNet builds on:</p> <ul> <li>Satorras, Hoogeboom, Welling, E(n) equivariant graph neural networks, ICML 2021 (the encoder).</li> <li>Radford et al., CLIP, ICML 2021 (the alignment objective).</li> <li>Batatia et al., MACE-MP-0 (stability screening).</li> </ul>"},{"location":"start/colab.html","title":"Run in Google Colab","text":"<p>Each notebook is a complete, runnable workflow. Training runs in your Colab session (free GPU is enough): your data never leaves your Google account.</p> Notebook What it does Time 01 \u2014 Quickstart published model, pick a target, read the report, explore the whole design space in 3D 5 min, CPU 02 \u2014 Your own data upload table + CIFs \u2192 check \u2192 train \u2192 generate \u2192 download everything 15 min + training 03 \u2014 Add a rule write a custom constraint in a few lines and see its effect before and after 5 min 04 \u2014 MEIDNet Studio in Colab the live Studio inside the notebook, then the same blocks driven from Python 10 min"},{"location":"start/colab.html#block-by-block-like-the-studio","title":"Block by block, like the Studio","text":"<p>Every notebook follows the seven blocks of the Studio, in order and in the same colours: Block 1 \u00b7 Data (blue), Block 2 \u00b7 Model (orange), Block 3 \u00b7 Family (green), Block 4 \u00b7 Rules (amber), Block 5 \u00b7 Targets (pink), Block 6 \u00b7 Search (dark green) and Block 7 \u00b7 Candidates (violet). Each block starts with a coloured header saying what it does, has a form for its settings, shows its result as charts (and crystals in 3D where it helps), and ends by printing what flows on to the next block.</p> <p>Run the cells from top to bottom (Runtime \u2192 Run all also works). Change a value in a form and re-run from that block on: the later blocks pick up the change.</p> <p>The notebooks are generated from <code>scripts/make_notebooks.py</code>, so they stay in step with the command line.</p>"},{"location":"start/quickstart.html","title":"5-minute quickstart","text":"In the browser (nothing to install)On your computerIn Colab <p>Open the live Studio. It runs the published perovskite model: move rule limits and targets and watch what passes, run a latent search and read each candidate's checklist, export the <code>meidnet.yaml</code>. (Search budgets are capped on the shared server; the local install below has no limits and takes your own data.)</p> <pre><code>pip install meidnet          # PyTorch CPU wheels are fine; CUDA is optional\nmeidnet demo                 # halide perovskites, band gap 2.0 eV \u2192 runs/demo/\n</code></pre> <p><code>demo</code> downloads the 2.8 MB published checkpoint, searches for 3 candidates (about a minute on a CPU), checks them against the family's rules and opens <code>generation_report.html</code>.</p> <pre><code>meidnet demo --family oxide --band-gap 3.0 --enthalpy -0.2 -n 5\nmeidnet studio               # the live workbench with the same model\n</code></pre> <p>Notebook 01 \u2014 quickstart</p>"},{"location":"start/quickstart.html#what-the-demo-did","title":"What the demo did","text":"<ol> <li>Targets \u2192 latent. The property encoder maps (band gap, enthalpy) to a point in the shared latent space.</li> <li>Search. A population of latents near that point is optimised so that the decoder's output matches the    targets and the family's soft terms (cubic cell, sensible charges, \u2026).</li> <li>Decode + rules. Each latent is decoded into one element per site; the composition is placed on the    family's prototype and must pass every hard rule (charge balance, tolerance factor, \u2026).</li> <li>Report. The best new, unique candidates are saved as CIFs with a card that shows each rule's value and    a gauge of predicted vs target property.</li> </ol>"},{"location":"start/quickstart.html#next","title":"Next","text":"<ul> <li>What data do I need? then Bring your own dataset.</li> <li>MEIDNet Studio to see the effect of every change live.</li> </ul>"},{"location":"start/what-data.html","title":"What data do I need?","text":""},{"location":"start/what-data.html#the-checklist","title":"The checklist","text":"You need Format Notes Structures CIF text in a <code>cif</code> column of the table, or one <code>&lt;id&gt;.cif</code> file per material in a folder up to <code>max_sites</code> atoms per cell (default 20); any elements Properties one or more numeric columns any units; each property is standardised unless you say otherwise An id a column whose values match the CIF file names (if you use files) <code>12</code>, <code>12.0</code> and <code>\"12\"</code> are treated as the same id A family the prototype your structures share (e.g. cubic ABX\u2083) needed for generation; training without one is possible (<code>align_to_prototype: false</code>) <p>A table like this is enough:</p> <pre><code>material_id,band_gap,dielectric,cif\nmat_001,1.42,18.3,\"# generated using pymatgen\ndata_CsPbBr3\n_cell_length_a 5.87 ...\"\nmat_002,2.16,11.7,\"...\"\n</code></pre> <p>CSV, Excel (<code>.xlsx</code>) and JSON tables are read; Parquet too once <code>pyarrow</code> is installed (<code>pip install \"meidnet[parquet]\"</code>).</p>"},{"location":"start/what-data.html#how-much-data","title":"How much data?","text":"<p>The published model used 11k structures. A few thousand is typical; a few hundred lets you try the pipeline, but expect rough predictions (the training report tells you how rough).</p>"},{"location":"start/what-data.html#what-happens-to-each-row","title":"What happens to each row","text":"<p><code>meidnet check</code> reads every row exactly as training would and writes <code>check_report.html</code> with one line per reason a row was skipped (unreadable CIF, too many atoms, missing value, does not match the prototype) and example ids, so every skipped row is accounted for.</p>"},{"location":"start/what-data.html#properties-and-ranges","title":"Properties and ranges","text":"<p>MEIDNet can only design for property values it has seen. The check report draws each property's distribution; the generation report flags targets outside the training range as extrapolation.</p>"},{"location":"start/what-data.html#which-family","title":"Which family?","text":"<ul> <li>Cubic ABX\u2083 perovskite \u2192 built-in <code>perovskite_abx3</code> (variants: oxide, halide, chalcogenide, nitride).</li> <li>Rock-salt-ordered double perovskite A\u2082BB\u2032X\u2086 \u2192 built-in <code>double_perovskite_a2bbx6</code>.</li> <li>Anything else with a fixed arrangement of sites \u2192 write a family file   (positions, which elements may occupy which site, their charges, a cell-size rule, rules).</li> <li>Mixed prototypes \u2192 you can still train (<code>align_to_prototype: false</code>) and use the model to predict/screen, but   generation needs one prototype.</li> </ul>"},{"location":"understand/how-it-works.html","title":"How MEIDNet works","text":""},{"location":"understand/how-it-works.html#one-latent-space-for-two-modalities","title":"One latent space for two modalities","text":"<pre><code>crystal (CIF) \u2500\u2500\u25ba SE(3)-equivariant encoder \u2500\u2500\u25ba z_c \u2500\u2510                     \u250c\u2500\u25ba crystal decoder   \u2500\u2500\u25ba lattice, species, positions\n                                                      \u251c\u2500\u2500 shared space \u2500\u2500\u2500\u2500\u2500\u2524\nproperties \u2500\u2500\u2500\u2500\u25ba property encoder (MLP) \u2500\u2500\u2500\u2500\u2500\u25ba z_p \u2500\u2518  pulled together     \u2514\u2500\u25ba property decoder  \u2500\u2500\u25ba properties\n                                                        (InfoNCE)\n</code></pre> <ul> <li>The crystal encoder embeds each atom's element, passes messages along neighbour bonds using only squared   distances (so rotating or translating the cell changes nothing), pools the atoms and adds the lattice.</li> <li>The property encoder maps the vector of (normalised) properties to the same space.</li> <li>Both are projected and normalised to unit length; a contrastive loss makes each structure's latent closest   to its own property latent and far from the others in the batch. Its weight ramps up during training so the   decoders stabilise first.</li> <li>The decoders are trained from the joint latent and from the property latent alone \u2014 that second   term is what makes inverse design possible: a point chosen from properties must decode into a structure.</li> </ul> <p>The training objective is</p> \\[ L = w_1 L_\\text{recon}(z_\\text{joint}) + w_2 L_\\text{prop}(z_\\text{joint}) + w_3 L_\\text{recon}(z_p) + w_4 L_\\text{prop}(z_p) + \\lambda(t)\\, L_\\text{InfoNCE}(z_c, z_p). \\]"},{"location":"understand/how-it-works.html#inverse-design-search-in-that-space","title":"Inverse design = search in that space","text":"<p>For each target:</p> <ol> <li>Start a population of latents at the property encoder's output for the target, plus a target-dependent    random direction (so different targets start apart) and noise.</li> <li>Optimise them by gradient descent on: distance of the decoded properties to the target (each objective    with its <code>loss</code> and <code>weight</code>); a cosine anchor to the target latent; the family's search terms (atoms    apart, cubic cell, positions near the prototype, expected charge zero, same element on equivalent sites,    loose tolerance factor); and repulsion between population members and from earlier successes (diversity).</li> <li>Decode: for each latent, choose one element per site group from the decoder's scores with temperature    and top-k sampling; the last group is drawn only from elements whose charge balances the others. Place the    elements on the prototype; the cell edge comes from ionic radii (<code>a = 2(r_B + r_X)</code> for perovskites).</li> <li>Rules: symmetry-refine, then apply every hard constraint. Only compositions that pass all are kept.</li> <li>Rank by distance to the target (<code>select_weight</code>, <code>select_loss</code>), skip duplicates and near-duplicate    latents, save as CIF.</li> </ol>"},{"location":"understand/how-it-works.html#what-family-means","title":"What \"family\" means","text":"<p>A family is data, not code: a prototype (site positions), site groups (allowed elements and oxidation states), a cell-size rule, soft search terms, and hard rules. The published perovskite tables are <code>families/perovskite_abx3.yaml</code>; the same machinery runs <code>double_perovskite_a2bbx6.yaml</code> and any file you write.</p>"},{"location":"understand/how-it-works.html#the-design-space","title":"The design space","text":"<p>For a prototype family, the set of compositions the decoder can output is finite. <code>meidnet space</code> and the Studio enumerate it and score each composition with the structure encoder. The search's candidates are a subset of this space; comparing the two views (encoder prediction vs search's decoded prediction) is a useful sanity check that the reports encourage.</p>"},{"location":"understand/how-it-works.html#reproducibility","title":"Reproducibility","text":"<p>MEIDNet 2 reproduces MEIDNet 1 exactly: the same featurisation, forward pass, training step and generation output (<code>tests/test_v1_regression.py</code> compares against the frozen v1 code in <code>tests/legacy_v1/</code>). One fix: v1's results depended on Python's hash seed through set iteration order; v2 fixes element order, so the same seed gives the same CIFs in any process.</p>"},{"location":"understand/interpretability.html","title":"Interpreting a candidate","text":"<p>MEIDNet shows its reasoning rather than a score. Where to look:</p>"},{"location":"understand/interpretability.html#before-trusting-the-model","title":"Before trusting the model","text":"<ul> <li>Training report \u2192 accuracy table. A property rated weak means its predictions are not much better than   guessing the mean; targets on it are unreliable.</li> <li>Training report \u2192 alignment. Low retrieval accuracy means property latents do not land near the right   structures; inverse design from them is guesswork.</li> <li>Check report \u2192 distributions. A target in a region with almost no training data is extrapolation.</li> </ul>"},{"location":"understand/interpretability.html#for-each-candidate","title":"For each candidate","text":"<ul> <li>The checklist \u2014 every rule's measured value and allowed window. These are deterministic chemistry   (charges, radii, distances), not model outputs.</li> <li>The gauges \u2014 where the prediction sits relative to the target and to the training range.</li> <li>Warnings \u2014 predicted value outside the training range (extrapolation) and latent hit the search limit   (the search pushed a latent to the <code>z_clip</code> bound; predictions there are less reliable). The model is trained   on latents of length 1 and every search moves them further out (typically to length 3\u20135), so the report says   this once for the whole run rather than on every card; each candidate's length is the <code>latent_norm</code> column of   <code>candidates.csv</code>.</li> <li>Encoder vs search \u2014 in the Studio, compare the search's decoded prediction with the structure encoder's   prediction for the same composition. Large disagreement is a red flag.</li> </ul>"},{"location":"understand/interpretability.html#for-the-run-as-a-whole","title":"For the run as a whole","text":"<ul> <li>The funnel \u2014 which rule removed most attempts. If it is a chemical rule (charge balance, tolerance   factor), the family's element lists may be too broad for the target; if it is no charge-balancing element,   the oxidation-state tables may be missing an entry.</li> <li>Rejected examples \u2014 what the search wanted to make; often more informative than what it kept.</li> <li>Latent map \u2014 whether the population spread out (diversity) or collapsed.</li> </ul>"},{"location":"understand/interpretability.html#what-a-candidate-is","title":"What a candidate is","text":"<p>A candidate is a composition on an ideal prototype with model-predicted properties. Whether it is stable, synthesisable, or has those properties is for <code>meidnet screen</code>, DFT and the lab to decide.</p>"},{"location":"understand/limits.html","title":"Scope and roadmap","text":""},{"location":"understand/limits.html#current-capabilities","title":"Current capabilities","text":"Structures any CIF with up to <code>max_sites</code> atoms (default 20); any elements Properties any number of scalar columns Training your data, your properties; the published settings by default Generation prototype families: a fixed arrangement of sites whose occupants and cell size are chosen. Built in: cubic ABX\u2083, A\u2082BB\u2032X\u2086; write your own in YAML Rules charge balance, tolerance and octahedral factors, bond windows, minimum distance, predicted-property windows, element filters, and your own Python rules Explanations reports for data, training and generation; the Studio for live what-if"},{"location":"understand/limits.html#not-yet-supported","title":"Not yet supported","text":"<ul> <li>New atomic arrangements. The decoder's positions are not used to create geometry; candidates are   compositions placed on the family prototype. Generating free geometry needs a decoder trained to produce it   (the published decoder receives the true positions during training).</li> <li>Spectra, images, text as modalities. The property modality is a vector of scalars. Binned XRD or DOS   could be added as a vector modality with its own encoder; it is designed for but not implemented.</li> <li>Large cells. <code>max_sites</code> can be raised, but memory and time grow with the square of it, and the   published model was trained on 5-atom cells.</li> <li>Fine-tuning the published model on new properties. Its property head has two outputs; train a new model.</li> </ul>"},{"location":"understand/limits.html#roadmap-in-order-of-likely-value","title":"Roadmap (in order of likely value)","text":"<ol> <li>Vector modalities (XRD, DOS) with a shared-latent encoder each.</li> <li>A <code>training.init_from</code> option to continue training from a checkpoint with the same properties.</li> <li>A geometry-generating decoder variant (trained from prototype positions instead of true positions \u2014    <code>model.decoder_coordinate_input: prototype</code> is the first step and is available as an experimental setting).</li> <li>More built-in families (spinel, Heusler, rocksalt/zincblende, MXene M\u2082XT\u2082).</li> </ol> <p>Contributions of family files and rules are the easiest way to help; see the GitHub repository.</p>"},{"location":"use/config.html","title":"The configuration file","text":"<p><code>meidnet.yaml</code> holds everything a run depends on. One file = one reproducible experiment: it is copied into every output folder and stored inside <code>model.pt</code>.</p> <pre><code>name: my_oxides                 # outputs \u2192 runs/my_oxides/\nplugins: []                     # Python files with custom rules (optional)\n\ndata:        # where the structures and properties come from        \u2192 meidnet check\nmodel:       # size of the network (defaults are the paper's)\ntraining:    # epochs, batch size, learning rate, alignment schedule \u2192 meidnet train\ngeneration:  # family, objectives, targets, rules overrides, budget  \u2192 meidnet generate\n</code></pre> <p>Rules of thumb:</p> <ul> <li>Unknown keys are errors. A typo such as <code>epocs:</code> stops the run with a message naming the key.</li> <li>Paths are relative to the YAML file, so a project folder can be moved or shared.</li> <li>Defaults reproduce the paper. Leave a section out and the published settings apply.</li> <li><code>meidnet schema</code> prints the JSON Schema (editors such as VS Code use it for completion and validation).</li> </ul> <p>The complete list of settings with their meaning and defaults is in the configuration reference.</p>"},{"location":"use/config.html#the-four-sections-in-one-picture","title":"The four sections in one picture","text":"<pre><code>data:\n  table: materials.csv            # one row per material\n  properties: [{column: band_gap, unit: eV}, {column: dielectric}]\n  align_to_prototype: true        # atoms re-ordered to the family prototype\n\ntraining:\n  epochs: 200\n  contrastive_weight: 5.0         # how strongly structure and property latents are pulled together\n\ngeneration:\n  family: perovskite_abx3\n  variant: oxide\n  objectives:                     # what to steer, and how to compare prediction with target\n    - {property: band_gap, loss: l2, weight: 10000}\n  targets:                        # one search per entry\n    - {band_gap: 2.0}\n  overrides:                      # change a family rule's parameters without editing the family file\n    tolerance_factor: {min: 0.85, max: 1.0}\n  exclude_elements: [Pb]\n</code></pre>"},{"location":"use/explore-3d.html","title":"Explore in 3D","text":"<p>The Studio can show any set of structures as a map linked to a 3D viewer: each point on the map is one crystal; click a point and its structure appears next to the map, where you can rotate and zoom it. The viewer is chemiscope, an open-source structure\u2013property explorer.</p> <p>Open it from any button marked \ud83e\uddca:</p> Button Where Shows \ud83e\uddca Explore all compositions in 3D Family block the design space: every composition of the family on its prototype \ud83e\uddca Map + 3D of all compositions Targets block the same design space, to compare predictions with your targets \ud83e\uddca All candidates in 3D Candidates block the candidates of your last search \ud83e\uddca See these structures in 3D Data block your data after Check my data, or the Perov-5 training data behind the published model \ud83e\uddca show in the 3D explorer the details window of any composition the design space, with that composition selected <p>You can also open the page directly in the explorer with a deep link: <code>?explore=space</code>, <code>?explore=candidates</code> or <code>?explore=data</code>.</p>"},{"location":"use/explore-3d.html#reading-the-map","title":"Reading the map","text":"<ul> <li>Axes. For the design space, the two predicted properties (for the published model: formation enthalpy   across, band gap up). For your data, your property columns. For candidates, their predicted properties. Any   other property can be put on an axis from the map's settings menu.</li> <li>Colour (design space): yellow = passes every rule, purple = rejected by at least one rule. Symbol:   pass / rejected. For candidates the colour is the search score (lower = closer to the targets).</li> <li>Properties you can plot or colour by:</li> </ul> Name Meaning <code>formula</code> the composition <code>site_A</code>, <code>site_B</code>, \u2026 the element on each site group <code>lattice_a</code> the cubic cell edge estimated from ionic radii (\u00c5) <code>pred_&lt;property&gt;</code> the model's prediction from the structure encoder, with its unit <code>rule_&lt;name&gt;</code> the value each rule measures, e.g. <code>rule_tolerance_factor</code> <code>passes_all</code>, <code>verdict</code> 1 / \"pass\" when every rule passes (with the limits currently set on the page) <code>score</code>, <code>round</code>, <code>target_&lt;property&gt;</code>, <code>still_passes</code> candidates only: search score, round found, the target, and whether it still passes the current rules <p>The design space shows up to 2,000 compositions: those passing every rule first, then a fixed random sample of the rest. Your data shows the structures that passed Check my data; the Perov-5 training data shows its first 1,500 structures.</p>"},{"location":"use/explore-3d.html#clicking-a-point","title":"Clicking a point","text":"<p>The crystal appears on the right: drag to rotate, scroll to zoom. The panel below lists every property of that structure. Bonds and the unit cell are drawn. Design-space structures are the ideal prototype filled with that composition's elements, at the cell size the family's lattice rule estimates.</p>"},{"location":"use/explore-3d.html#when-chemiscope-cannot-load","title":"When chemiscope cannot load","text":"<p>The explorer is downloaded from a public CDN the first time you open it (about 3 MB). Without internet access, or if the CDN is blocked, the Studio switches to its built-in viewer: a scatter plot of the same two axes (coloured points pass every rule) linked to the Studio's own crystal viewer \u2014 click a point, drag the cell to rotate, double-click to spin. Nothing else changes.</p>"},{"location":"use/explore-3d.html#save-the-dataset","title":"Save the dataset","text":"<p>\u2b07 dataset JSON saves what you see as a chemiscope dataset. Open it later at chemiscope.org (the site reads the file inside your browser; nothing is uploaded), share it with a colleague, or put it in a paper's supplementary material.</p> <p>The same datasets can be produced without the Studio:</p> <pre><code>meidnet space meidnet.yaml --chemiscope space.json   # the design space of your configuration\n</code></pre> <p>and from Python:</p> <pre><code>from meidnet.studio.chemiscope import space_dataset, candidates_dataset, records_dataset\n</code></pre> <p>In the Colab notebooks, the <code>show3d(...)</code> helper displays structures inline with <code>chemiscope.show</code>.</p>"},{"location":"use/explore-3d.html#credits","title":"Credits","text":"<p>chemiscope is a third-party library (BSD-3-Clause licence) by G. Fraux, R. K. Cersonsky and M. Ceriotti. If you use the 3D explorer in published work, please cite: G. Fraux, R. K. Cersonsky, M. Ceriotti, Chemiscope: interactive structure-property explorer for materials and molecules, Journal of Open Source Software 5, 2117 (2020).</p>"},{"location":"use/reports.html","title":"Reading the reports","text":"<p>Each command writes a self-contained HTML file in <code>runs/&lt;name&gt;/</code>. All three open with a one-sentence verdict (green / amber / red), show the evidence, and end with what to do next.</p>"},{"location":"use/reports.html#check_reporthtml","title":"<code>check_report.html</code>","text":"<ul> <li>Verdict: how many rows are usable; the most common problem if many were skipped.</li> <li>What MEIDNet read: table, columns, structure source, family and how many structures were re-ordered to   the prototype.</li> <li>Rows that were skipped: one row per reason with counts and example ids, and a \"how do I fix these?\" box.</li> <li>Your properties: histogram, min/median/max, and warnings (e.g. \"72 % of band-gap values are 0 \u2014 the model   sees few examples of other values\").</li> <li>Your structures: atoms per cell, most common elements.</li> </ul>"},{"location":"use/reports.html#training_reporthtml","title":"<code>training_report.html</code>","text":"<ul> <li>Verdict: are the properties predicted well from structure alone?</li> <li>How accurate: per property, typical error (MAE), natural spread (std), R\u00b2, and a word \u2014 good if the error   is under a quarter of the spread, fair under a half, weak otherwise \u2014 with predicted-vs-true plots.</li> <li>Did the two modalities align: retrieval accuracy (a structure's own property vector is its nearest match   x % of the time) and the cosine curve. Inverse design depends on this.</li> <li>Training curves and notes (e.g. the alignment ramp longer than training).</li> </ul>"},{"location":"use/reports.html#generation_reporthtml","title":"<code>generation_report.html</code>","text":"<ul> <li>Verdict: how many candidates, and whether any predicted value lies outside the training range.</li> <li>Per target: the funnel (compositions alive after each rule \u2014 the biggest drop is the rule that limits the   family most), the candidate cards, and examples of rejected compositions with the rule and the   measured value.</li> <li>A candidate card shows: elements per site, cell edge, round found, a gauge per objective (training range   as a track, target as a tick, prediction as a dot), the checklist with each rule's value and allowed   window (hover for the explanation), and warnings: prediction outside the training range, or a latent far   from where training latents live (predictions less reliable).</li> <li>Where the search went: latents before and after optimisation projected to two dimensions.</li> </ul>"},{"location":"use/reports.html#candidatescsv-and-generationjson","title":"<code>candidates.csv</code> and <code>generation.json</code>","text":"<p>The CSV has one row per candidate with targets, predictions, elements, cell edge, every rule's value, score and file name. The JSON holds the full funnel counts, rejected examples and settings \u2014 everything the report shows.</p>"},{"location":"use/studio.html","title":"MEIDNet Studio \u2014 the live workbench","text":"<pre><code>meidnet studio                 # the published perovskite model\nmeidnet studio meidnet.yaml    # your project\n</code></pre> <p>A page opens in your browser. The workflow is drawn as a strip of seven coloured blocks; click a block to open it. Whatever you change, the blocks after it update at once, and a line under the strip says, in words, what the change did to each later block \u2014 the way a LabVIEW diagram shows data flowing after you turn a knob.</p> <p>No installation? Use the hosted Studio \u2014 the same page, running on a shared server.</p> <pre><code> Data \u2500\u2500\u25ba Model \u2500\u2500\u25ba Family \u2500\u2500\u25ba Rules \u2500\u2500\u25ba Targets \u2500\u2500\u25ba Search \u2500\u2500\u25ba Candidates\n 11,356    MAE       halide     5 rules   Eg 2.0     idle       0 found\n materials 0.17 eV   924 comps  27 pass   \u0394Hf \u22120.1\n</code></pre>"},{"location":"use/studio.html#reading-the-workflow-strip","title":"Reading the workflow strip","text":"<p>Each block has its own colour, used everywhere on the page (its panel, its charts, its chips in the flow chain):</p> Block Colour What it holds Data blue the table of materials the model learns from Model orange the trained network (published or yours) Family green the crystal prototype and the elements allowed on each site Rules amber the hard checks every candidate must pass Targets pink the property values you want Search dark green the latent search of the paper Candidates violet what the search found <ul> <li>The dot in the corner of a block tells its state: grey = idle, green = fine, amber = just changed or needs a   look (for example a target outside the training range), red = error, pulsing = running.</li> <li>The two lines inside a block are its live status: \"924 comps\", \"27 pass all\", \"MAE 0.17\", \"running\".</li> <li>The label on each arrow says what flows to the next block: \"11,356 materials \u00b7 2 properties\",   \"924 compositions\", \"27 pass all rules\", \"3 found\".</li> <li>Moving dots on the arrows appear after every change: they travel from the block you edited to every block   it reaches, and those blocks flash. This is the quickest way to see what depends on what.</li> </ul> <p>Dashed pink lines are always a target; a violet \u2605 is always something the search found; green \u2713 means passes.</p>"},{"location":"use/studio.html#the-flow-chain-and-the-flow-log","title":"The flow chain and the flow log","text":"<p>Under the strip, the flow chain repeats the last change in words, block by block, for example:</p> <p>Rules: Goldschmidt tolerance factor max 1.05 \u2192 1 \u279c Rules: pass all rules 27 \u2192 23 (\u22124) \u279c Targets: best match CsSnBr\u2083 \u2192 CsGeBr\u2083 \u279c Candidates: 1 of 3 found candidates would now be rejected</p> <p>Click a step to open that block.</p> <p>The Flow log below it keeps the last 60 changes. Click an entry to replay its flow on the strip.</p>"},{"location":"use/studio.html#what-is-live-and-how","title":"What is live, and how","text":"<p>For a prototype family the set of possible compositions is finite (924 for halide perovskites, 10,340 for halide double perovskites). When you open a family the server builds this design space once: every composition on the prototype, the value of each rule's descriptor (tolerance factor, octahedral factor, \u2026) and the model's predicted properties from the structure encoder. From then on, the page can answer instantly:</p> You change You see immediately a rule's limit how many compositions still pass that rule and all rules; the histogram of the descriptor with the window shaded; the funnel an element (click a chip to exclude it) compositions removed, new counts, new best matches a target the compositions predicted closest; the target drawn on the predicted-property histogram and scatter; an extrapolation warning if it lies outside the training range the comparison kind (as close as possible / at most / at least) the ranking anything, after a search which of the found candidates would now be rejected <p>Descriptor values are computed by the same Python functions that <code>meidnet generate</code> uses, so what the page shows is what the search enforces.</p>"},{"location":"use/studio.html#the-two-directions-of-the-model","title":"The two directions of the model","text":"<p>The live tables use structure \u2192 property (encode the prototype structure, decode its properties): fast, and it scores every composition. The Search block runs the paper's property \u2192 structure latent optimisation and shows its progress (target, round, step, loss), the log, and each candidate as it is found, with its checklist. Both views are of the same network; comparing them tells you whether the search found what the encoder considers best, and whether the predictions agree.</p>"},{"location":"use/studio.html#the-blocks","title":"The blocks","text":"<ul> <li>Data \u2014 how many materials were usable, why rows were skipped, property distributions, the elements   present; Bring your own data (below).</li> <li>Model \u2014 properties, units, training ranges, validation errors, loss and alignment curves; Train your own   model (below).</li> <li>Family \u2014 prototype sites with a rotatable 3D preview of the cell, element chips per site group with their   charges (click to exclude), cell-size rule.</li> <li>Rules \u2014 each rule with its explanation, parameters, pass count and descriptor histogram; add a   predicted-property window rule; the funnel.</li> <li>Targets \u2014 one card per objective: target value, comparison kind, weights; the predicted-property scatter   of all compositions (coloured = passes, grey = rejected, \u2605 = found by the search). Click any point or table row   for its checklist and its crystal in 3D.</li> <li>Search \u2014 budget (candidates, latents per round, rounds, steps), a rough time estimate, Run/Stop, progress,   log, link to the full report.</li> <li>Candidates \u2014 one card per candidate: a rotatable 3D cell, elements, gauges of predicted values against the   targets, the rule checklist, warnings, CIF links.</li> </ul> <p>Crystals are drawn by a small built-in viewer: drag to rotate, double-click to spin.</p>"},{"location":"use/studio.html#beginner-view-and-researcher-mode","title":"Beginner view and Researcher mode","text":"<p>The page opens in the beginner view: every block has an Input \u279c Logic \u279c Output strip with live numbers (\"924 compositions \u279c 5 hard rules, applied in order \u279c 27 pass every rule\") and the name of the Python function that does the work, a short In short explanation, and a glossary in the side panel (composition, rule, target, latent space, extrapolation).</p> <p>The Researcher mode switch in the header trades these for the technical view: the \u270e Edit as text YAML drawer, and a Behind the scenes panel in every block with the <code>meidnet.yaml</code> and the Python of that block. The choice is remembered by your browser.</p>"},{"location":"use/studio.html#behind-the-scenes","title":"Behind the scenes","text":"<p>At the bottom of every block, Behind the scenes shows the same settings as the part of <code>meidnet.yaml</code> they come from and as the Python that does the same thing \u2014 with the values currently on the page. Copy either into your own project; nothing on the page is hidden from the command line.</p>"},{"location":"use/studio.html#bring-your-own-data","title":"Bring your own data","text":"<p>The Data block has an adapter: it maps your column names to what MEIDNet expects. Nothing in your files is changed.</p> <ol> <li>Upload a table \u2014 CSV, Excel (<code>.xlsx</code>) or JSON, one row per material \u2014 with an id column, numeric property    columns and the structures, either as a column of CIF text or as a zip of <code>&lt;id&gt;.cif</code> files (select both files    together). Parquet works too when <code>pyarrow</code> is installed.</li> <li>Map the columns. The Studio suggests the id column, the CIF column and the numeric property columns; pick    the properties you want, give each a unit (eV, eV/atom, \u2026) and optionally a label, and choose the family (and    variant) your structures belong to. Align to prototype re-orders each structure's atoms onto the family's    sites, which generation needs; raise the prototype tolerance if many rows are skipped as \"does not match the    prototype\".</li> <li>Check my data. Every row is parsed and checked; you see how many are usable, why the others were skipped    (with examples), the distribution of each property and the elements present.</li> <li>Train in the Model block: choose the number of epochs and press \u25b6 Train. The loss and alignment curves    grow while it trains; at the end the validation errors appear and every block switches to your model and    your properties \u2014 the design space is re-scored, the targets are your properties. Each property's typical    error (MAE) is judged in words against the spread (standard deviation) of that property in your data \u2014 good below a quarter of    the spread, fair below half, weak above \u2014 the same rule the training report uses; \"matched correctly\" is    how often a validation structure is matched to its own properties.</li> <li>Switch between your model and the published model at any time in the Model block. Start over    removes your upload and model from the session.</li> </ol> <p>A few epochs give a rough model in a minute or two; 100\u2013200 epochs are typical for real work \u2014 run <code>meidnet train</code> locally (with a GPU if you have one) and open the result with <code>meidnet studio meidnet.yaml</code>.</p> <p>On the hosted Studio</p> <p>Each browser tab has its own private session: other visitors never see your data, model or candidates, and a session's files are deleted after an hour without activity. The shared machine allows up to 1,500 rows and 8 MB per upload, 30 training epochs, 20 atoms per cell, and two jobs (training or search) at a time across all visitors; search budgets are capped too. Run the Studio on your own computer to lift every limit.</p>"},{"location":"use/studio.html#edit-the-configuration-as-text","title":"Edit the configuration as text","text":"<p>\u270e Edit as text in the header opens a drawer with the <code>generation</code> section of <code>meidnet.yaml</code> exactly as the blocks have it: family, variant, excluded elements, rule overrides, extra rules, objectives, targets and search budget. Change anything and press Apply \u2014 the blocks update and the flow chain shows what changed in each.</p> <ul> <li>Validate only checks without applying. Errors name the exact setting, for example   <code>steps: Input should be greater than or equal to 1</code>.</li> <li>On the hosted Studio, values above the shared limits are accepted but capped, and a note says so:   <code>rounds 99 \u2192 6 (limit on this shared server)</code>.</li> <li>Predicted-property windows added in the Rules block appear as <code>extra_constraints</code>, so the file round-trips:   apply it, download it, run it with <code>meidnet generate</code>.</li> </ul> <p>\u2b07 Download meidnet.yaml (also in the Search block) writes the whole configuration \u2014 rules switched off on the page are listed under <code>disabled_rules</code>, excluded elements become <code>exclude_elements</code>, changed limits become <code>overrides</code> \u2014 so the command line reproduces exactly what the page shows. Settings the blocks do not show (for example search terms set in your own <code>meidnet.yaml</code>) are kept as they are.</p>"},{"location":"use/studio.html#explore-in-3d","title":"Explore in 3D","text":"<p>Buttons marked \ud83e\uddca open the 3D explorer: a map with one point per structure, linked to a 3D viewer \u2014 the design space of the family, the candidates of a search, or the training data. See Explore in 3D.</p>"},{"location":"use/studio.html#deep-links","title":"Deep links","text":"<p>Add to the address to open the page at a given place:</p> Address ends with Opens <code>?panel=rules</code> the Rules block (any block: <code>data</code>, <code>model</code>, <code>family</code>, <code>rules</code>, <code>targets</code>, <code>search</code>, <code>candidates</code>) <code>?explore=space</code> the 3D explorer on the design space (<code>candidates</code> or <code>data</code> work too) <p>Both can be combined: <code>?panel=targets&amp;explore=space</code>.</p>"},{"location":"use/studio.html#dark-mode","title":"Dark mode","text":"<p>The \ud83c\udf19 / \u2600\ufe0f button in the header switches between the light and the dark theme; your browser remembers the choice.</p>"},{"location":"use/studio.html#static-copy-for-a-website","title":"Static copy for a website","text":"<pre><code>meidnet studio --export-static docs/studio.html\n</code></pre> <p>embeds the state and design spaces into one HTML file. Everything works without a server except uploading, training and running a new search \u2014 useful for a plain web host. The hosted Studio is the full server instead:</p> <pre><code>meidnet studio --public --host 0.0.0.0 --port 7860 --run-root /tmp/runs --docs-dir site\n</code></pre> <p><code>--public</code> gives every visitor a private session and caps the budgets so the shared machine stays responsive; <code>--docs-dir</code> also serves a built copy of this documentation at <code>/docs/</code>.</p>"},{"location":"use/your-data.html","title":"Bring your own dataset","text":"<p>This page takes a table of structures and properties to designed candidates. Every step writes a report that explains what happened; nothing requires Python.</p>"},{"location":"use/your-data.html#0-install","title":"0. Install","text":"<pre><code>pip install meidnet\n</code></pre>"},{"location":"use/your-data.html#1-describe-your-data-meidnet-init","title":"1. Describe your data: <code>meidnet init</code>","text":"<pre><code>meidnet init --name my_oxides --table materials.csv \\\n             --id-column material_id --cif-column cif \\\n             --properties band_gap dielectric \\\n             --family perovskite_abx3 --variant oxide\n</code></pre> <p>This writes a commented <code>meidnet.yaml</code>. Open it: every line says what it does. If your CIFs are files instead of a column, set <code>cif_column: null</code> and <code>structures_dir: structures/</code>.</p> <pre><code>data:\n  table: materials.csv\n  id_column: material_id\n  cif_column: cif\n  properties:\n    - {column: band_gap,   unit: eV}\n    - {column: dielectric, unit: \"\"}\n  val_fraction: 0.1\n  max_sites: 20\n  align_to_prototype: true\n</code></pre>"},{"location":"use/your-data.html#2-is-it-usable-meidnet-check","title":"2. Is it usable? <code>meidnet check</code>","text":"<pre><code>meidnet check meidnet.yaml\n</code></pre> <pre><code>412 of 430 materials usable.\n  skipped     14  does not match the perovskite_abx3 prototype   e.g. m_031, m_118, m_204\n  skipped      4  missing property value                          e.g. m_077, m_300\nreport: runs/my_oxides/check_report.html\n</code></pre> <p>The report shows the verdict, the distribution of each property, the elements present and, for each skip reason, how to fix it (for example raise <code>data.prototype_tolerance</code> for distorted cells).</p>"},{"location":"use/your-data.html#3-learn-the-latent-space-meidnet-train","title":"3. Learn the latent space: <code>meidnet train</code>","text":"<pre><code>meidnet train meidnet.yaml            # 200 epochs by default\nmeidnet train meidnet.yaml --epochs 5 # a quick smoke test first\n</code></pre> <p><code>training_report.html</code> answers three questions: how accurate is each property prediction (error vs natural spread \u2192 good / fair / weak), did the two modalities align (retrieval accuracy), and what were the training curves. The model is saved as <code>runs/my_oxides/model.pt</code> with the property names, units and normalisation inside; <code>meidnet info runs/my_oxides/model.pt</code> describes it.</p> <p>Time</p> <p>About 21 s per epoch per 11k structures on a laptop GPU; roughly 10\u00d7 slower on CPU. Use Colab's free GPU (notebook 02) if you have none.</p>"},{"location":"use/your-data.html#4-design-meidnet-generate","title":"4. Design: <code>meidnet generate</code>","text":"<p>Set the targets in <code>meidnet.yaml</code>:</p> <pre><code>generation:\n  family: perovskite_abx3\n  variant: oxide\n  objectives:\n    - {property: band_gap,   loss: l2,       weight: 10000}\n    - {property: dielectric, loss: at_least, weight: 5000}\n  targets:\n    - {band_gap: 2.0, dielectric: 20}\n    - {band_gap: 3.0, dielectric: 20}\n  per_target: 4\n  exclude_elements: [Pb, Cd]\n</code></pre> <pre><code>meidnet generate meidnet.yaml\n</code></pre> <p><code>generation_report.html</code> shows, per target, the funnel (how many attempts each rule removed), one card per candidate (elements, cell, each rule's measured value and window, predicted vs target gauges, warnings about extrapolation) and examples of rejected compositions. CIFs and <code>candidates.csv</code> are in <code>runs/my_oxides/generation/</code>.</p>"},{"location":"use/your-data.html#5-explore-live-meidnet-studio","title":"5. Explore live: <code>meidnet studio</code>","text":"<pre><code>meidnet studio meidnet.yaml\n</code></pre> <p>Change a rule's limit, exclude an element or move a target and see immediately how many compositions pass and which are predicted closest; run searches and export the YAML. Studio guide.</p>"},{"location":"use/your-data.html#6-confirm","title":"6. Confirm","text":"<p>Predicted properties are estimates. <code>meidnet screen runs/my_oxides/generation</code> (MACE universal potential, optional extra) gives a first stability filter; DFT or experiment has the last word.</p>"}]}
\ No newline at end of file
diff --git a/site/start/colab.html b/site/start/colab.html
index 6db16ac0c650013b81308904d9b35efb8e5c1ffa..1dc3aa374b19cd2ffda2f7eaae1cf212e60523b2 100644
--- a/site/start/colab.html
+++ b/site/start/colab.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1670,7 +1670,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/start/quickstart.html b/site/start/quickstart.html
index 1d4634b370af37695230d62cc93b39686691d3e0..e875edc6f6bcc1bd050aeaed37021ca73b079e8a 100644
--- a/site/start/quickstart.html
+++ b/site/start/quickstart.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1329,10 +1329,10 @@
     <ul class="md-nav__list" data-md-component="toc" data-md-scrollfix>
       
         <li class="md-nav__item">
-  <a href="#what-you-just-saw" class="md-nav__link">
+  <a href="#what-the-demo-did" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What you just saw
+        What the demo did
       
     </span>
   </a>
@@ -1681,7 +1681,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2099,10 +2099,10 @@
     <ul class="md-nav__list" data-md-component="toc" data-md-scrollfix>
       
         <li class="md-nav__item">
-  <a href="#what-you-just-saw" class="md-nav__link">
+  <a href="#what-the-demo-did" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What you just saw
+        What the demo did
       
     </span>
   </a>
@@ -2164,7 +2164,7 @@ meidnet<span class="w"> </span>studio<span class="w">               </span><span
 </div>
 </div>
 </div>
-<h2 id="what-you-just-saw">What you just saw<a class="headerlink" href="#what-you-just-saw" title="Permanent link">&para;</a></h2>
+<h2 id="what-the-demo-did">What the demo did<a class="headerlink" href="#what-the-demo-did" title="Permanent link">&para;</a></h2>
 <ol>
 <li><strong>Targets → latent.</strong> The property encoder maps (band gap, enthalpy) to a point in the shared latent space.</li>
 <li><strong>Search.</strong> A population of latents near that point is optimised so that the decoder's output matches the
diff --git a/site/start/what-data.html b/site/start/what-data.html
index a951ecddc0c08a4bcb41cfbd2d4db190a74f33b3..4caba064feabbc5cc4c3565efe20d5754c53780e 100644
--- a/site/start/what-data.html
+++ b/site/start/what-data.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1714,7 +1714,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2254,7 +2254,7 @@ but expect rough predictions (the training report tells you how rough).</p>
 <h2 id="what-happens-to-each-row">What happens to each row<a class="headerlink" href="#what-happens-to-each-row" title="Permanent link">&para;</a></h2>
 <p><code>meidnet check</code> reads every row exactly as training would and writes <code>check_report.html</code> with one line per
 reason a row was skipped (unreadable CIF, too many atoms, missing value, does not match the prototype) and
-example ids. Nothing is silently dropped.</p>
+example ids, so every skipped row is accounted for.</p>
 <h2 id="properties-and-ranges">Properties and ranges<a class="headerlink" href="#properties-and-ranges" title="Permanent link">&para;</a></h2>
 <p>MEIDNet can only design for property values it has seen. The check report draws each property's distribution;
 the generation report flags targets outside the training range as extrapolation.</p>
diff --git a/site/studio.html b/site/studio.html
index 358f73899f47ddbd5bce176ddf76e55e2a2068d6..47586ebccc917aa97eb4006c3ec33bedba4b5852 100644
--- a/site/studio.html
+++ b/site/studio.html
@@ -198,7 +198,7 @@ body>div:empty:not([id]):not([class]){position:absolute;top:0;left:0;width:0;hei
     </div>
   </section>
   <div class="card strip"><div class="strip-scroll"><svg id="pipe" viewBox="0 0 1270 150" aria-label="MEIDNet workflow"></svg></div></div>
-  <div class="chain" id="chain"><span class="lead">Change anything above or in a block — what it does to the blocks after it appears here.</span></div>
+  <div class="chain" id="chain"><span class="lead">Changes made above or in a block are listed here, with their effect on the blocks after it.</span></div>
   <details id="flowlog"><summary>Flow log (0)</summary><ol id="flowlog-list"></ol></details>
   <div class="cols">
     <div id="panel"><div class="card loading-card"><h2>Loading MEIDNet Studio…</h2><p class="muted">Fetching the model and every composition of the perovskite family. This takes a few seconds; on the hosted Studio the very first visit can take up to a minute while the server wakes up.</p><div class="prog"><i class="indeterminate"></i></div></div></div>
@@ -331,7 +331,7 @@ async function boot(){
   if(BLOCKS.some(b=>b.id===qs.get('panel'))) S.panel = qs.get('panel');
   await loadFamily();
   render();
-  chainText(`Ready: ${nfmt(S.ev.rows.length)} compositions, ${nfmt(S.ev.alive.length)} pass every rule. Change anything — its effect on every later block appears here.`);
+  chainText(`Ready: ${nfmt(S.ev.rows.length)} compositions, ${nfmt(S.ev.alive.length)} pass every rule. Changes and their effect on the later blocks are listed here.`);
   if(!STATIC) await resumeJobs();
   if(['space','candidates','data'].includes(qs.get('explore'))) openExplore(qs.get('explore'));
   api('/api/data').then(d=>{ S.data=d; S.dataOwn = !!(S.state.session && S.state.session.checked); statusAll(); updateEdges(); render(); }).catch(()=>{});
@@ -652,7 +652,7 @@ function render(){
   const b=BLOCKS.find(x=>x.id===S.panel); const n=BLOCKS.indexOf(b)+1;
   P.dataset.block=S.panel;
   P.innerHTML = `<div class="card"><div class="eyebrow">Block ${n} of 7 · ${esc(b.name)}</div><h2>${esc(b.name)}</h2><p class="purpose">${esc(b.purpose)}</p>
-    <div class="plain beginner-only"><b>In plain words.</b> ${esc(b.plain)}</div>${ioStrip(S.panel)}${f?f():''}${behind(S.panel)}</div>`;
+    <div class="plain beginner-only"><b>In short.</b> ${esc(b.plain)}</div>${ioStrip(S.panel)}${f?f():''}${behind(S.panel)}</div>`;
   document.querySelectorAll('.node').forEach(nd=>nd.classList.toggle('sel', nd.id==='node-'+S.panel));
   if(!S.ev && S.fam) recompute(null, null); else { statusAll(); updateEdges(); renderLive(); }
   renderLegend(); mountAll3d();
@@ -860,7 +860,7 @@ function trainBox(){
         <tr><td>typical error (MAE)</td>${Object.entries(p.val.mae).map(([k,v])=>`<td>${fmt(v)}${esc(unit(k))}</td>`).join('')}<td>${(100*p.val.retrieval_top1).toFixed(0)}%</td></tr>
         <tr><td>quality</td>${q.map(x=>`<td class="${x.css}">${x.word}</td>`).join('')}<td></td></tr></table>
         <p class="small muted beginner-only">The typical error is compared with how much each property varies in your data: under 25% of that spread is <b>good</b>, under 50% <b>fair</b>, otherwise <b>weak</b>. "Matched correctly" is how often a material's structure and its properties land closest to each other in the model's hidden space (higher is better). A few epochs give a rough model; more epochs or more data usually help.</p>`; }
-    if(st.done&&!st.error) b+=`<div class="verdict good">✅ Model ready — the blocks below now use <b>your</b> model and properties.${st.report?` <a href="/files/${esc(st.report)}" target="_blank">training report</a> · <a href="/files/${esc(st.report.replace('training_report.html','model.pt'))}" download="model.pt">download model.pt</a>`:''}</div>`; }
+    if(st.done&&!st.error) b+=`<div class="verdict good">Model ready. The blocks below now use <b>your</b> model and properties.${st.report?` <a href="/files/${esc(st.report)}" target="_blank">training report</a> · <a href="/files/${esc(st.report.replace('training_report.html','model.pt'))}" download="model.pt">download model.pt</a>`:''}</div>`; }
   return b;
 }
 async function startTrain(){
@@ -913,13 +913,13 @@ function panelFamily(){
   b+=`<h3>Elements per site group <span class="muted small">(click to exclude / include)</span></h3>`;
   for(const [g,grp] of Object.entries(f.groups)){ b+=`<h4>${esc(g)} — ${esc(grp.description)} · ${grp.slots.length} site${grp.slots.length>1?'s':''}</h4><div>`;
     for(const el of grp.sample){ const off=S.ui.exclude.has(el); const ox=(grp.oxidation_states[el]||[]).map(q=>(q>0?'+':'')+q).join('/'); b+=`<span class="chip ${off?'off':''}" onclick="toggleEl(${jsq(el)})" title="charges ${esc(ox)}"><span style="display:inline-block;width:9px;height:9px;border-radius:50%;background:${CPK[el]||'#999'};border:1px solid var(--line);margin-right:5px;vertical-align:-1px"></span>${esc(el)}<small class="muted"> ${esc(ox)}</small></span>`; } b+='</div>'; }
-  b+=`<div class="help beginner-only">Excluding an element removes every composition containing it — watch the Family → Rules numbers change. To change which elements a family allows at all, copy <code>${esc(S.ui.family)}.yaml</code> and edit its <code>groups:</code>.</div>`;
+  b+=`<div class="help beginner-only">Excluding an element removes every composition that contains it; the Family and Rules counts update accordingly. To change which elements a family allows at all, copy <code>${esc(S.ui.family)}.yaml</code> and edit its <code>groups:</code>.</div>`;
   return b;
 }
 
 /* ── Rules ── */
 function panelRules(){
-  const ev=S.ev||evaluateSpace(); let b=`<p class="muted small">Move a limit and watch the count change — the search enforces exactly these windows.</p>`;
+  const ev=S.ev||evaluateSpace(); let b=`<p class="muted small">Moving a limit updates the count; the search applies exactly these windows.</p>`;
   for(const c of S.fam.constraints){ const key=ruleKey(c); const r=S.ui.rules[key]; const d=S.fam.space.rows.map(x=>x.d[key]).filter(Number.isFinite);
     b+=`<div class="rule ${r.enabled?'':'off'}"><div class="hd"><input type="checkbox" ${r.enabled?'checked':''} onchange="toggleRule(${jsq(key)},this.checked)"><b>${esc(ruleTitle(key))}</b><span class="muted small"><b style="color:var(--accent-text)">${nfmt(ev.perRule[key]??0)}</b> of ${nfmt(ev.rows.length)} pass</span></div><p class="small muted">${esc(ruleText(key))}</p>`;
     const num=Object.entries(r.params).filter(([k,v])=>typeof v==='number' && !['cutoff'].includes(k));
@@ -1005,7 +1005,7 @@ function renderLive(){
   L.innerHTML=b;
 }
 function renderLegend(){ const beg=document.body.classList.contains('beginner');
-  $('legend').innerHTML = beg ? `<h4>Glossary</h4><dl class="gloss"><dt>Composition</dt><dd>one element per site of the prototype, e.g. CsPbBr₃.</dd><dt>Rule</dt><dd>a hard check with a measured value and an allowed window.</dd><dt>Target</dt><dd>a property value you want; the model predicts how close each composition gets.</dd><dt>Latent space</dt><dd>the hidden coordinates where structures and properties meet; the search moves through it.</dd><dt>Extrapolation</dt><dd>a target outside the training range — predictions there are guesses.</dd></dl><p class="small muted">Switch on Researcher mode in the header for the configuration as text and the Python behind each block.</p>`
+  $('legend').innerHTML = beg ? `<h4>Glossary</h4><dl class="gloss"><dt>Composition</dt><dd>one element per site of the prototype, e.g. CsPbBr₃.</dd><dt>Rule</dt><dd>a hard check with a measured value and an allowed window.</dd><dt>Target</dt><dd>a property value you want; the model predicts how close each composition gets.</dd><dt>Latent space</dt><dd>the hidden coordinates where structures and properties meet; the search moves through it.</dd><dt>Extrapolation</dt><dd>a target outside the training range; predictions there are less reliable.</dd></dl><p class="small muted">Switch on Researcher mode in the header for the configuration as text and the Python behind each block.</p>`
     : `<h4>How to use</h4><ol class="small" style="padding-left:18px;margin:4px 0"><li><b>Family</b>: pick the crystal type; exclude elements.</li><li><b>Rules</b>: move limits; counts update live.</li><li><b>Targets</b>: set values; see the closest compositions.</li><li><b>Search</b> runs the latent optimisation; <b>Candidates</b> explains each result.</li><li><b>Data/Model</b>: upload your table, train, and design with your own properties.</li></ol>`; }
 
 /* ═══════════════════════════ actions ═══════════════════════════ */
diff --git a/site/understand/how-it-works.html b/site/understand/how-it-works.html
index 6ddfa42a79ebc7be2f069d270e5dc706cd6dd8f9..924a7c40d523fd8f567c1c7db4803e0b2cca6e2f 100644
--- a/site/understand/how-it-works.html
+++ b/site/understand/how-it-works.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1714,7 +1714,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/understand/interpretability.html b/site/understand/interpretability.html
index db5296363acab956ff51e9ef79912d3275c71f1a..d4a6ef90dc4a8a532b0f83a6ca376c695104b648 100644
--- a/site/understand/interpretability.html
+++ b/site/understand/interpretability.html
@@ -27,7 +27,7 @@
     
     
       
-        <title>Why you can trust (or not) a candidate - MEIDNet Prism</title>
+        <title>Interpreting a candidate - MEIDNet Prism</title>
       
     
     
@@ -80,7 +80,7 @@
     <div data-md-component="skip">
       
         
-        <a href="#why-you-can-trust-or-not-a-candidate" class="md-skip">
+        <a href="#interpreting-a-candidate" class="md-skip">
           Skip to content
         </a>
       
@@ -113,7 +113,7 @@
         <div class="md-header__topic" data-md-component="header-topic">
           <span class="md-ellipsis">
             
-              Why you can trust (or not) a candidate
+              Interpreting a candidate
             
           </span>
         </div>
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1621,7 +1621,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -1639,7 +1639,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -1698,10 +1698,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#what-meidnet-never-claims" class="md-nav__link">
+  <a href="#what-a-candidate-is" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What MEIDNet never claims
+        What a candidate is
       
     </span>
   </a>
@@ -2154,10 +2154,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#what-meidnet-never-claims" class="md-nav__link">
+  <a href="#what-a-candidate-is" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What MEIDNet never claims
+        What a candidate is
       
     </span>
   </a>
@@ -2184,7 +2184,7 @@
   
 
 
-<h1 id="why-you-can-trust-or-not-a-candidate">Why you can trust (or not) a candidate<a class="headerlink" href="#why-you-can-trust-or-not-a-candidate" title="Permanent link">&para;</a></h1>
+<h1 id="interpreting-a-candidate">Interpreting a candidate<a class="headerlink" href="#interpreting-a-candidate" title="Permanent link">&para;</a></h1>
 <p>MEIDNet shows its reasoning rather than a score. Where to look:</p>
 <h2 id="before-trusting-the-model">Before trusting the model<a class="headerlink" href="#before-trusting-the-model" title="Permanent link">&para;</a></h2>
 <ul>
@@ -2215,7 +2215,7 @@
 <li><strong>Rejected examples</strong> — what the search <em>wanted</em> to make; often more informative than what it kept.</li>
 <li><strong>Latent map</strong> — whether the population spread out (diversity) or collapsed.</li>
 </ul>
-<h2 id="what-meidnet-never-claims">What MEIDNet never claims<a class="headerlink" href="#what-meidnet-never-claims" title="Permanent link">&para;</a></h2>
+<h2 id="what-a-candidate-is">What a candidate is<a class="headerlink" href="#what-a-candidate-is" title="Permanent link">&para;</a></h2>
 <p>A candidate is a <em>composition on an ideal prototype with model-predicted properties</em>. Whether it is stable,
 synthesisable, or has those properties is for <code>meidnet screen</code>, DFT and the lab to decide.</p>
 
diff --git a/site/understand/limits.html b/site/understand/limits.html
index 0cd4f2397c925112992bf284720ccba5cc607e92..d8125e8a50d48a451fe85e606fbd5adb5854a5b2 100644
--- a/site/understand/limits.html
+++ b/site/understand/limits.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1612,7 +1612,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -1693,10 +1693,10 @@
     <ul class="md-nav__list" data-md-component="toc" data-md-scrollfix>
       
         <li class="md-nav__item">
-  <a href="#what-works-today" class="md-nav__link">
+  <a href="#current-capabilities" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What works today
+        Current capabilities
       
     </span>
   </a>
@@ -1704,10 +1704,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#what-does-not-work-yet" class="md-nav__link">
+  <a href="#not-yet-supported" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What does not work (yet)
+        Not yet supported
       
     </span>
   </a>
@@ -2110,10 +2110,10 @@
     <ul class="md-nav__list" data-md-component="toc" data-md-scrollfix>
       
         <li class="md-nav__item">
-  <a href="#what-works-today" class="md-nav__link">
+  <a href="#current-capabilities" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What works today
+        Current capabilities
       
     </span>
   </a>
@@ -2121,10 +2121,10 @@
 </li>
       
         <li class="md-nav__item">
-  <a href="#what-does-not-work-yet" class="md-nav__link">
+  <a href="#not-yet-supported" class="md-nav__link">
     <span class="md-ellipsis">
       
-        What does not work (yet)
+        Not yet supported
       
     </span>
   </a>
@@ -2163,7 +2163,7 @@
 
 
 <h1 id="scope-and-roadmap">Scope and roadmap<a class="headerlink" href="#scope-and-roadmap" title="Permanent link">&para;</a></h1>
-<h2 id="what-works-today">What works today<a class="headerlink" href="#what-works-today" title="Permanent link">&para;</a></h2>
+<h2 id="current-capabilities">Current capabilities<a class="headerlink" href="#current-capabilities" title="Permanent link">&para;</a></h2>
 <table>
 <thead>
 <tr>
@@ -2198,7 +2198,7 @@
 </tr>
 </tbody>
 </table>
-<h2 id="what-does-not-work-yet">What does not work (yet)<a class="headerlink" href="#what-does-not-work-yet" title="Permanent link">&para;</a></h2>
+<h2 id="not-yet-supported">Not yet supported<a class="headerlink" href="#not-yet-supported" title="Permanent link">&para;</a></h2>
 <ul>
 <li><strong>New atomic arrangements.</strong> The decoder's positions are not used to create geometry; candidates are
   compositions placed on the family prototype. Generating free geometry needs a decoder trained to produce it
diff --git a/site/use/config.html b/site/use/config.html
index 6e5436ac766227221b3a2736ae05f9c3d4a46bf0..dfcaa103b6f55a29f6869319caf9964d9b1fb9f1 100644
--- a/site/use/config.html
+++ b/site/use/config.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1670,7 +1670,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/use/explore-3d.html b/site/use/explore-3d.html
index 38e7dec45236e49df70f8d7625f4b7c80b1acc26..7d864cf595cbc2be5958a98fce351a6cf0056824 100644
--- a/site/use/explore-3d.html
+++ b/site/use/explore-3d.html
@@ -704,7 +704,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -1005,7 +1005,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1714,7 +1714,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/use/reports.html b/site/use/reports.html
index dafa6445781cb7d93498998a2f315eeaf4289726..c732de335909672039954a881abf21dabe6c2b0f 100644
--- a/site/use/reports.html
+++ b/site/use/reports.html
@@ -704,7 +704,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -994,7 +994,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1703,7 +1703,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
diff --git a/site/use/studio.html b/site/use/studio.html
index ee19996bf92e9ae6b92b816d6a0ff7ccbfab873f..d4b0410a4d30e4816a7886bfc23f03d79f9a61ac 100644
--- a/site/use/studio.html
+++ b/site/use/studio.html
@@ -27,7 +27,7 @@
     
     
       
-        <title>MEIDNet Studio (live workbench) - MEIDNet Prism</title>
+        <title>MEIDNet Studio - MEIDNet Prism</title>
       
     
     
@@ -113,7 +113,7 @@
         <div class="md-header__topic" data-md-component="header-topic">
           <span class="md-ellipsis">
             
-              MEIDNet Studio (live workbench)
+              MEIDNet Studio
             
           </span>
         </div>
@@ -713,7 +713,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -731,7 +731,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -1093,7 +1093,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1802,7 +1802,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2526,7 +2526,7 @@ the encoder considers best, and whether the predictions agree.</p>
 <h2 id="beginner-view-and-researcher-mode">Beginner view and Researcher mode<a class="headerlink" href="#beginner-view-and-researcher-mode" title="Permanent link">&para;</a></h2>
 <p>The page opens in the <strong>beginner view</strong>: every block has an <strong>Input ➜ Logic ➜ Output</strong> strip with live numbers
 ("924 compositions ➜ 5 hard rules, applied in order ➜ 27 pass every rule") and the name of the Python function that
-does the work, a short <strong>In plain words</strong> explanation, and a <strong>glossary</strong> in the side panel (composition, rule,
+does the work, a short <strong>In short</strong> explanation, and a <strong>glossary</strong> in the side panel (composition, rule,
 target, latent space, extrapolation).</p>
 <p>The <strong>Researcher mode</strong> switch in the header trades these for the technical view: the <strong>✎ Edit as text</strong> YAML
 drawer, and a <strong>Behind the scenes</strong> panel in every block with the <code>meidnet.yaml</code> and the Python of that block.
diff --git a/site/use/your-data.html b/site/use/your-data.html
index 995763b445371d94e0b54a1fb2c3b2ee89a1ef2b..f3e65ffa650682fdd10a907f55a593ffcdf14f6e 100644
--- a/site/use/your-data.html
+++ b/site/use/your-data.html
@@ -699,7 +699,7 @@
   <span class="md-ellipsis">
     
   
-    MEIDNet Studio (live workbench)
+    MEIDNet Studio
   
 
     
@@ -898,7 +898,7 @@
   <span class="md-ellipsis">
     
   
-    Capabilities (what works today)
+    Capabilities
   
 
     
@@ -1736,7 +1736,7 @@
   <span class="md-ellipsis">
     
   
-    Why you can trust (or not) a candidate
+    Interpreting a candidate
   
 
     
@@ -2289,7 +2289,7 @@ reason, how to fix it (for example raise <code>data.prototype_tolerance</code> f
 <div class="highlight"><pre><span></span><code>meidnet<span class="w"> </span>train<span class="w"> </span>meidnet.yaml<span class="w">            </span><span class="c1"># 200 epochs by default</span>
 meidnet<span class="w"> </span>train<span class="w"> </span>meidnet.yaml<span class="w"> </span>--epochs<span class="w"> </span><span class="m">5</span><span class="w"> </span><span class="c1"># a quick smoke test first</span>
 </code></pre></div>
-<p><code>training_report.html</code> answers three questions in plain words: <em>how accurate is each property prediction</em>
+<p><code>training_report.html</code> answers three questions: <em>how accurate is each property prediction</em>
 (error vs natural spread → good / fair / weak), <em>did the two modalities align</em> (retrieval accuracy), and <em>what
 were the training curves</em>. The model is saved as <code>runs/my_oxides/model.pt</code> with the property names, units and
 normalisation inside; <code>meidnet info runs/my_oxides/model.pt</code> describes it.</p>