Selvauma/eas-mlip-training-data

Engineered Atomic Structures (EAS) — Training Data for Complex MLIPs

The training dataset behind EAS-MLIP: a foundation machine-learned interatomic potential for structures engineered from distinct fundamental units — crystals, surfaces, molecules, electrolytes, monomers, polymers, and their interfaces — rather than one fixed material class. This repo is the data; the model itself (currently training, 4-seed MACE ensemble) will follow as a separate release once it converges.

Scale

Split	Frames	Fraction
train	24,585,711	80%
validation	3,287,191	10%
test	3,095,248	10%
Total	30,968,150

Drawn from 45,440 source calculations (mostly VASP OUTCAR trajectories, plus ASE-trajectory and pre-processed DeepMD/extxyz sources) across 930 distinct fidelity groups — different DFT settings/pseudopotentials/codes are tracked as separate fidelities, not silently pooled together.

Source-calculation QC states (from the audit that built this set): usable: 8,446 · partial: 31,053 · incomplete: 5,542 · failed: 399 (failed/incomplete excluded from the splits above).

Files

File	Size	Content
data/train.extxyz	220 GB	training split
data/validation.extxyz	30 GB	validation split
data/test.extxyz	28 GB	held-out test split
data/splits.parquet	4.1 GB	split manifest (geometry/trajectory/fidelity IDs → split assignment)

~282 GB total. Every entry is standard extended XYZ — readable by ASE, MACE, DeepMD's converters, etc. out of the box. Example record:

52
Lattice="9.8589 0.0 0.0 0.0 9.8589 0.0 0.0 0.0 9.8589" Properties=species:S:1:pos:R:3:REF_forces:R:3 REF_energy=-167.43632685 eas_geometry_hash=14a0bcbe... eas_source_id=b49bf6a5... eas_fidelity_id=vasp-9da35442b51c eas_trajectory_id=b49bf6a5...:block-000001 eas_ionic_step=100 pbc="T T T"
Li       4.68819000       5.01980000       4.53632000       0.07217600       0.11163300      -0.02517600
...

Beyond the standard REF_energy/REF_forces labels, every frame carries full provenance: eas_geometry_hash (dedup key), eas_source_id + eas_trajectory_id (which calculation/trajectory it came from, down to the eas_ionic_step), and eas_fidelity_id (which DFT setting group). Nothing is anonymous — any frame can be traced back to its source calculation.

Split methodology — leakage-checked

Splits are group-aware by eas_trajectory_id, not per-frame random — consecutive ionic steps from the same MD/relaxation trajectory always land in the same split, so the model is never validated on near-duplicate frames of something it trained on. This was checked, not assumed: the split manifest's own leakage report (leaking_geometries: [], leaking_trajectory_groups: [], valid: true, seed 20260804) confirms zero cross-split leakage at both the geometry and trajectory-group level.

Why "Engineered Atomic Structures"

Most MLIP training sets are scoped to one material family. This one is built across deliberately heterogeneous, purpose-built structures — the "atomic structures" are engineered for specific interfaces, defects, and compositions relevant to battery-materials research, not sampled from one crystal's phase space. The multi-fidelity tracking (930 groups) and leakage-resistant splitting exist specifically because pooling heterogeneous, unequal-quality sources without that bookkeeping silently corrupts benchmarks.

Uploading this (a note on scale)

This is not a same-day huggingface-cli upload like a 100 MB model file — ~282 GB needs the resumable large-file path:

pip install -U huggingface_hub
huggingface-cli login
huggingface-cli repo create eas-mlip-training-data --type dataset -y

python -c "
from huggingface_hub import upload_large_folder
upload_large_folder(
    repo_id='Selvauma/eas-mlip-training-data',
    repo_type='dataset',
    folder_path='.',
)
"

upload_large_folder chunks, parallelizes, and resumes on failure — plain upload_folder/upload_file are not built for this size and will be slow/fragile. Expect this to take a while regardless (hours, depending on your outbound bandwidth from this cluster) — run it in tmux/screen or as a background/batch job, not in a terminal you might close.

If you want a lighter starting point instead of the full 282 GB, a deduplicated/filtered subset already exists on disk at /projects/nmclps/EAS_MLIP/campaigns/full-v0.1/direct/direct_train_filtered.extxyz (7.8 GB) — ask and I'll stage that as a -sample companion repo.

Citation

Selva Chandrasekaran Selvaraj, University of Illinois Chicago. Produced by the EAS-MLIP data pipeline (github.com/selvachandrasekaranselvaraj/EAS_MLIP), orchestrated on HPC via HPCA (github.com/selvachandrasekaranselvaraj/hpca).