junlinw/ai4sci-atomistic-materials

Crystal Potential Learning

One repository for the training data, starting model, trained checkpoints, physical evaluation inputs/reference labels and measured results for AI4Sci task PR #49. The task learns one energy/force/stress potential and evaluates crystal stability, phonons and elasticity. It complements CrystalGen's structure-generation task.

Contents

Directory	Contents	Role
training/	500,000 MPTrj+sAlex training frames; 5,000 labeled-dev frames	Fit weights / select checkpoints; each split is 50% MPTrj and 50% sAlex
models/uma-s-1p1.pt	Original UMA-S-1.1 checkpoint	Fixed starting backbone for reproducing both trials
models/3ep/	Exported 3-epoch model, energy audit, export metadata and training manifest	Reference starter result; selected step 46,875
models/30ep-selected/	Exported checkpoint selected by labeled-dev loss	Step 400,000 / epoch 25.6 from the completed 30-epoch run
models/30ep-final/	Exported final epoch-30 checkpoint	Step 468,750; secondary comparison
evaluation/val/	1,200 geometries and reference labels	1,000 WBM discovery + 100 MDR phonon + 100 MP elasticity cases
evaluation/test/	6,000 geometries and reference labels	5,000 WBM + 500 phonon + 500 elasticity; reserved final grading
evaluation/regression/	28 known-failure geometries and manifest	Diagnostic only
splits/v1/, panels/v1/	Frozen split assignments, exclusions and panel IDs	Provenance and independent reconstruction
records/	Both trials' measured curves, metrics, predictions, timings and comparison figures	Reproduce the reported comparison
licenses/, LICENSES.md	Source terms, attribution and redistribution notices	Apply per component; not one blanket license

manifest.json inventories every materializable payload, its component, output path, compressed/unpacked size and SHA-256. models/index.json identifies all four checkpoints. The original v1 training files and split bytes are unchanged; v2 adds checkpoints, prepared physical panels and the 30-epoch records. Use an immutable HF revision.

Download selected components

From the updated task directory in PR #49:

# Development preparation; no final-test files.
uv run --frozen python environment/data/materialize.py \
  --components training protocol initial-model eval-val
# Reproduce reported predictions from the trained reference models.
uv run --frozen python environment/data/materialize.py \
  --components trained-models records
# Final evaluator preparation only; this command does not run evaluation.
uv run --frozen python environment/data/materialize.py --components eval-test

The downloader includes component license notices automatically. To download via HF without the task downloader, use snapshot_download(repo_id="junlinw/ai4sci-atomistic-materials", repo_type="dataset", revision="<pinned commit>", allow_patterns=[...]) and retain LICENSES.md and licenses/* with the selected payloads. JSONL files ending in .gz need decompression for the physical runner. Model artifacts use FAIRChem 2.16.0 / PyTorch 2.8.0; inference recipes, scoring environments and dependency locks are in the PR.

Model provenance

The original checkpoint is unchanged from facebook/UMA at revision f611b917d9c68566bbbeccbb0aa0f7cad1696cb2. Original UMA energy conventions differ from Materials Project. The adapted models use the pretrained backbone plus a new conservative E/F/S head, both trained on MP-compatible raw total energies. Use each adapted checkpoint with its adjacent energy-audit.json; do not apply its provenance to the original checkpoint. Exported weights are non-EMA and contain no optimizer state. The three- and 30-epoch trials began from the same original initialization with seed 41 and global batch 32; their cosine horizons differ.

All UMA files and derivatives remain under the FAIR Chemistry License and its Acceptable Use Policy, included verbatim in licenses/. Redistribution here does not relicense the models under the data licenses or imply Meta endorsement.

Evaluation separation

Test reference labels are included for the trusted evaluator, not for fitting, checkpoint selection, development feedback or prediction inputs. Download on the operator host and mount only permitted files into each runtime; never mount the whole repository into a model process. Geometry-only inputs and labels are separate. The 6,000-case test has not been evaluated. Publishing already-public reference values does not change that reservation or authorize evaluating it.

Training and labeled dev are composition-disjoint, with at most four frames per source-parent and no atom cap. All compositions/parent IDs from the original physical evaluation universe are excluded from adaptation. These exclusions do not establish independence from UMA pretraining. The physical panels are custom subsets, not an official leaderboard submission; counts across suites can include the same material.

Training rows contain IDs, periodic geometry, energy in eV, forces in eV/Å and ASE tensile-positive stress in eV/ų. MPTrj compression-positive kbar stresses were converted; released sAlex ASE units were retained. No MP2020 correction was applied to training energies. Prepared evaluation labels retain the task's pinned reference conventions; source pins are included in evaluation/sources.json.

Measured comparison

Metric	3 epochs	Long: selected epoch 25.6	Long: final epoch 30
Discovery F1 ↑	0.8563	0.8496	0.8571
Hull MAE, meV/atom ↓	20.448	19.735	19.809
Phonon max / mean frequency MAE, K ↓	53.682 / 19.767	49.686 / 22.128	45.594 / 20.499
Bulk / shear modulus MAE, GPa ↓	4.852 / 9.990	6.093 / 10.916	6.449 / 10.811

All checkpoints scored 1,200/1,200 validation cases. Training took 2 h 16 min 14 s versus 25 h 5 min 4 s on eight H100s. Physical validation took 14 min 22 s / 11 min 41 s / 10 min 57 s respectively. The final-checkpoint time is from the original full evaluation call; the saved raw summary was later overwritten by a partial cached replay timer. The timing audit preserves both results and explains the correction.

The longer run gives mixed physical results at about 11× training time. One seed; no claim of convergence, statistical significance or state-of-the-art performance.