You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These are embeddings derived from Meta's UMA (Universal Model for Atoms) weights and are therefore governed by the FAIR Chemistry License. By requesting access you agree to the FAIR Chemistry License and its Acceptable Use Policy, you confirm you are not located in and will not redistribute to China, Russia, Belarus, or any comprehensively sanctioned jurisdiction, and you agree to redistribute these Materials (or any derivative works) only under the same FAIR Chemistry License terms with a copy of the agreement provided.

Log in or Sign Up to review the conditions and access this dataset content.

Layer-wise frozen UMA-S representations for 154,875 Materials Project crystals — GATED

⚠️ Access-gated. These embeddings are derivative works of Meta's UMA weights and are redistributed under the FAIR Chemistry License, not CC-BY. You must accept the license to download. See Licensing below. The permissive half of this study (ORB embeddings + Materials Project metadata/targets + results) is in the companion dataset NatalijaStepurko/orb-mp-embeddings.

Precomputed UMA (FairChem) structure-level embeddings for the same 154,875 Materials Project crystals as the companion dataset, from the study Out-of-training properties in frozen interatomic potential embeddings are decodable and ordered by physical scale (arXiv:2611.21116). The study has two papers, at the AI4Mat workshop (AI for Accelerated Materials Design) and the AI for Science workshop, both at NeurIPS 2026; this dataset, its open companion and the code repository are the same for both.

Three levels — read this first

Start here embeddings/ — the reusable resource (needs the companion repo's data/materials.parquet for targets)
Quick inspection sample/ — the same 1,000 materials as the open repo
Paper reproduction only reproduction/ — not needed to use the embeddings

Layout

path what
embeddings/node/uma-node-{family}-{layer}.parquet 36 shards — 6 equivariant families × 6 layers
embeddings/edge/uma-edge-{family}-{layer}.parquet 30 shards — 6 families × 5 layers
manifest.parquet one row per shard: read-out, representation, layer, dimensionality, rows, and whether it is rotation-invariant — the distinction §3.4 turns on
reproduction/ inputs for the cross-model and rotation analyses, and the UMAP coordinates
sample/sample_1000_uma.parquet 1,000-material slice for inspection

Node layers: embed, block_0..3, backbone. Edge layers: embed, block_0..3. Families (equivariant channel decompositions of the 9×128 messages): node full (1152-d), ang_norm (384), l0 (128, invariant scalar), l1_polar, l2_polar, geom_norm; edge flat, ang_norm, l0, l1_norm, l2_norm, geom_norm.

Schema: mp_id, model("uma"), level, family, layer, emb (list<float32>). The 2-D UMAP coordinates that used to sit in every shard are fitted visualisation outputs, not part of the representation; they are kept as one optional file, reproduction/umap_coords_uma.parquet (mp_id, level, family, layer, umap_x, umap_y). One row per material per file (154,875), matching the ORB layout in the companion repo.

Coverage. 154,875 materials, matching the companion repo and the figure the paper reports. Join to the companion repo's data/materials.parquet on mp_id for targets, labels and the fixed splits.

Naming convention

One model's data carries that model in the name; files holding both models or neither do not. So the shards are uma-…, while reproduction/convergence_atom_sample/ holds both models' aligned vectors and is itself unprefixed, with orb-… and uma-… files inside it.

reproduction/

convergence_atom_sample/ — the 30,005 atom-aligned vectors behind the cross-model grids (the resampled confidence intervals of the peaks draw from a 45,002-atom pool that is not deposited and needs the per-atom caches, i.e. re-extraction) (the convergence grids, the stitching table and the supplementary results on cross-model convergence): ORB node 256-d and UMA ang-norm 384-d at every layer, row i the same atom in both, plus _index.parquet giving each row's element. The ORB half lives here because splitting the pair across repositories would destroy the alignment that makes it useful; placing ORB data behind the gate is over-restrictive, the reverse would not be permissible.

rotation_control_uma/ holds the exact pooled activations used for the paper's rotation-invariance control. They are not required for ordinary embedding reuse; they are provided so the control can be reproduced without re-running gated UMA inference. UMA features for the 1,000-material rotation control (the rotation-control table) at all 18 UMA configurations, in both states. state=rotated comes from a fresh forward pass after an independent random rotation and is stored nowhere else, so without it that table cannot be checked without the model weights. Rotation invariance is verifiable directly: the mean relative change between states is exactly 0 for ang_norm and l0, 0.05 for the raw full tensor and 2.7 for l1_polar. rotation_control_metadata.parquet gives the exact transform per material — the integer random_state and the 3×3 matrix — so a reader with the model weights can regenerate the rotated activations rather than trusting the shipped ones.

The sample

sample/sample_1000_uma.parquet covers the same 1,000 mp_ids as the companion repo's sample, drawn with seed 42 stratified by crystal system, so the open and gated slices join on mp_id. It holds UMA vectors only — Materials Project metadata and ORB vectors are in the open repo, and no UMA vector is written outside this gate.

Scope of this repo: embeddings only

This gated repo contains only the UMA-derived embeddings — nothing else is behind the gate. The two result tables that the released code checks a rerun against (every probe score and the bootstrap metric table, UMA rows included) are in the permissive companion dataset and need no license acceptance: they are aggregate research results, not redistributions of the Materials. Every other result is regenerated by the code from the embeddings here.

Practical consequence: you can read, cite and re-analyse every published number without requesting access. The gate applies only if you want the raw UMA embedding vectors themselves.

Not included (too large to host)

The full per-atom UMA node embeddings (~67 GB fp16). The 30,005-atom aligned subsample above covers every cross-model result the paper reports; the full set is available on request.

Reproducing the results

Join these UMA pooled embeddings to the companion materials table on mp_id (for targets

  • the fixed seed-42 split__<target> flags), then refit the same estimators used in the paper — see the runnable example in the companion dataset's card (identical recipe, swapping embeddings/node/orb-node-<layer>.parquet for a embeddings/node/uma-node-<family>-<layer>.parquet file; targets, splits and the structures are in the companion repository). This reproduces the UMA property-prediction and all-layers-concat results (pooled embeddings; verified for ORB to within solver noise, Δ ≤ 0.001).

Reproducibility scope. The atom-level CKA/SVCCA cross-model convergence is reproducible from reproduction/convergence_atom_sample/, on the same 30,005 atoms the paper used.

Licensing

These embeddings are derivative works of the UMA model weights and are governed by the FAIR Chemistry License (Meta / FAIR Chemistry):

  • Redistribution of the Materials or any derivative works is permitted only under the terms of the FAIR Chemistry License, and you must provide a copy of the agreement to recipients.
  • Use must comply with the FAIR Chemistry Acceptable Use Policy and applicable Trade Control Laws.
  • Not available in China, Russia, Belarus, or comprehensively sanctioned jurisdictions.
  • Commercial use is permitted under the license; you must acknowledge use of the Materials in any publication.

Full text: LICENSE-FAIR-Chemistry.txt (the Agreement) and USE_POLICY.md (the Acceptable Use Policy, incorporated into it by Section 1(b)(iii)) — both mirroring the UMA model repository, https://huggingface.co/facebook/UMA . The underlying Materials Project structures/targets remain CC-BY-4.0 (in the companion dataset).

Citation

  • UMA: Wood et al., UMA: A Family of Universal Models for Atoms (Meta FAIR Chemistry).
  • Materials Project: Jain et al., APL Materials 1, 011002 (2013).
  • This dataset / paper: Stepurko, N. and Laulainen, J. E. M. (2026). Out-of-training properties in frozen interatomic potential embeddings are decodable and ordered by physical scale. arXiv:2611.21116, doi:10.48550/arXiv.2611.21116. Presented at the AI4Mat workshop (AI for Accelerated Materials Design) and the AI for Science workshop, both at NeurIPS 2026. Code: https://github.com/Natalija-Stepurko/mlip-embedding-probes
@misc{stepurko2026outoftraining,
  title         = {Out-of-training properties in frozen interatomic potential embeddings are decodable and ordered by physical scale},
  author        = {Stepurko, Natalija and Laulainen, Joonatan E. M.},
  year          = {2026},
  eprint        = {2611.21116},
  archivePrefix = {arXiv},
  primaryClass  = {physics.chem-ph},
  doi           = {10.48550/arXiv.2611.21116},
  url           = {https://doi.org/10.48550/arXiv.2611.21116}
}
Downloads last month
422

Paper for NatalijaStepurko/uma-mp-embeddings-gated