METHODS.md

Methods

1. Generation

Structures come from CrystaLLM (a language model trained on CIF files), prompted with a composition. Most prompts are picked automatically ("autopilot"); a few were typed by a person. origin and prompt record which.

How the autopilot picks prompts changed on 2026-11-01 (prompt_mode records it from then on):

• before: 30% no composition (CrystaLLM chooses), 70% random cations and counts plus a random anion, with no charge balance. On our own hull data these random formulas were the worst source by far.
• since: 45% substitution (one element of a stable Alexandria compound swapped for a likely replacement: pymatgen's data-mined substitution probabilities for ions, same-group neighbours otherwise, metal-for-metal in intermetallics; formulas already in the 2.8M-structure training index removed; f-block lanthanides left out because plain PBE handles them poorly), 40% CrystaLLM chooses, 15% random but charge-balanced.

2. ML screen

Each structure is relaxed with MACE (float64). The verdict comes from the formation energy relative to the elements and from how much the cell changed in relaxation: rock solid / plausible / borderline / unstable / falls apart.

3. Novelty

Composition and structure are matched (pymatgen StructureMatcher) against a local index of known structures and, since 2026-10-28, live against Materials Project and Alexandria via OPTIMADE. known and rediscovered records include the matching ID. Before 2026-10-28 only the local index was used, which missed some known compounds; affected records were re-audited.

Since 2026-11-03 a discovery must also be a composition nobody has reported, in any structure:

• before generation, autopilot formulas (substitution and random modes) are skipped if they appear in a local list of known compositions (the 2024 index above + current Alexandria PBE/PBEsol + both GNoME releases + Materials Project, rebuilt weekly); "surprise" crystals, where CrystaLLM picks the formula, are checked right after generation and skipped before MACE. The README's prompt-mode table counts these skips.
• before cataloguing, viable crystals get the online prior-art check described in §7 (about 32 sources incl. OQMD, COD, MPDS, NOMAD, OpenAlex and arXiv). A hit gives novelty known composition and no catalogue entry.
• a new arrangement of a known formula (the old new structure status) is no longer catalogued; the 638 existing ones and every other catalogued record that failed the composition check were re-labelled known composition on 2026-11-03, with their old catalogue number and name kept in the novelty record.
• control prompts (5% of autopilot) ask for known stable compounds on purpose, to check that the generator and MACE still recognise real crystals. They are never catalogued.

4. ML convex hull

Nightly, catalogued structures get an energy above the hull: MACE energies of the candidate vs MACE-relaxed Alexandria phases near the hull in the same chemical system. decomposes_to lists the hull products and fractions. Values below about -0.15 eV/atom are flagged suspect (almost always an ML artefact).

5. DFT re-check

Candidates with ML e_hull ≤ 0.025 eV/atom and ≤ 30 atoms go to DFT, together with their decomposition products on Alexandria's PBE hull (≤ 64 atoms each).

• Code: Quantum ESPRESSO 6.7, PBE, SSSP 1.1.2 efficiency pseudopotentials; cutoffs = max over the elements of the SSSP-recommended wavefunction/charge cutoffs.
• Smearing: Marzari-Vanderbilt, 0.01 Ry. Spin: non-magnetic, except systems containing Fe/Co/Ni/Mn/Cr, which are run spin-polarised from a 2 μB start on those atoms.
• Each structure: vc-relax on a coarse k-grid (spacing 0.4 Å⁻¹; forces 1e-3 Ry/bohr, pressure 0.5 kbar), then an scf on a finer grid (0.25 Å⁻¹, conv_thr 1e-7). Energies come from the final scf only.
• Candidate and competitors are computed with identical settings, so the hull is internally consistent.
• Pass: e_hull ≤ 0.025 eV/atom. That margin is about the size of PBE's typical error and of entropy at room temperature; a pass means "worth a closer look", nothing more.
• Some jobs carry a note (e.g. a relaxation stopped by hand when it was clearly not going to pass, or an extra competitor checked afterwards). Read it before using the number.

6. Properties (passes only)

PBE band gap and density of states at the Fermi level; elastic tensor from ±1% strains in the six Voigt directions (ions relaxed at fixed cell), with Born stability check. PBE underestimates band gaps, often by a factor of ~2.

7. Phonons (passes only)

Since 2026-11-03 a pass first gets a composition check (dft/<id>/novelty.json, summarised in the prior_reports column of dft_results.csv): any entry with the same reduced formula, in any structure, in Materials Project, Alexandria, OQMD or COD, or the formula (every element ordering) in a paper title/abstract indexed by OpenAlex. Any hit means no phonon run unless reviewed. This is wider than the structure match in §3: #1856 SrCaCdP2 passed §3 but its structure is OQMD entry 5879242 (a computed prediction, no ICSD record).

phonopy finite displacements on a supercell with every side ≥ 7 Å, after a tight fixed-cell relaxation; frequencies on a 10×10×10 mesh. Any mode softer than -0.1 THz away from the Γ acoustic modes = unstable. The supercell is small, so long-wavelength instabilities can be missed.

Known limits

• Hulls only include competitors that exist in Alexandria (and, for the ML hull, phases near its hull). A missing competitor can make a candidate look more stable than it is.
• PBE alone; no van der Waals, no Hubbard U, no finite-temperature free energies.
• Pass ≠ synthesisable. No candidate here has been made in a lab.