Multi-Fidelity Training of Machine-Learned Force Fields — Dataset

Dataset accompanying the paper Understanding Multi-Fidelity Training of Machine-Learned Force-Fields.

File Structure

data/
├── data.lmdb # LMDB database with atomic structures and labels
├── metadata.parquet # Lightweight metadata index
├── schema.json # Schema for the metadata
├── {method}_train_{a,b,c,d}.json # Train/validation/test split definitions (12 files)

• data.lmdb — The main database containing atomic positions, atomic numbers, energies, and forces for each structure.
• metadata.parquet — A metadata index with columns: formula, conformation_idx, method, n_atoms, energy, forces_present, energy_unit, forces_unit, idx.
• {method}_train_{a,b,c,d}.json — Split files defining train, validation, and test indices for each method (dft, xtb, cc) and training group (a–d). Indices reference entries in the LMDB database.
• schema.json — Schema definition for the metadata fields.

Code

The code to reproduce the experiments in the paper is available at github.com/microsoft/multi-fidelity-training-mlff.

Citation

@online{Gardner2025Understanding,
 title = {Understanding Multi-Fidelity Training of Machine-Learned Force-Fields},
 author = {Gardner, John L. A. and Schulz, Hannes and Helie, Jean and Sun, Lixin and Simm, Gregor N. C.},
 date = {2025-06-17},
 eprint = {2506.14963},
 eprinttype = {arXiv},
 eprintclass = {physics},
 doi = {10.48550/arXiv.2506.14963},
 url = {http://arxiv.org/abs/2506.14963}
}

License

This dataset is released under the MIT License.

Contact

• John Gardner — johngardner@microsoft.com
• Gregor Simm — gregorsimm@microsoft.com