Datasets and Distillation Labels for the Paper "Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians"

We provide 6 data folders, which were used in our paper  Amin, I., Raja, S., Krishnapriyan, A.S. (2024). Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians. Accepted to ICLR 2025. arXiv:2501.09009.

md22_JMP_labels.tar.gz - md22 JMP (large and small, finetuned) Hessian Labels for Buckyball Catcher and Double Walled Nanotube splits

SPICE_MaceOFF_labels.tar.gz - SPICE Mace-OFF Hessian Labels

MPtrj_labels.tar.gz -  MPTrj Mace-MP Hessian Labels

spice_separated.tar.gz - SPICE subdatasets (lmdb) (Solvated Amino Acids, Molecules with Iodine, DES370K Monomers)

md22.tar.gz - MD22 datasets (lmdb) for buckyball catcher and double wall nanotube. Taken from the JMP repository (see paper).

MPtrj_separated_all_splits.zip -  MPtrj subdatasets (lmdb) filtered by property (Pm3m Spacegroup, Systems with Yttrium, Bandgap >= 5 meV). 

The original data was taken from the SPICE dataset , MPtrj dataset, and  md22 dataset

The repository for the paper, where these datasets can be used, is available at https://github.com/ASK-Berkeley/MLFF-distill.

If you found any of this useful, please consider citing the paper:

@article{amin2025distilling,      title={Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians},      author={Ishan Amin, Sanjeev Raja, and Krishnapriyan, A.S.},      journal={International Conference on Learning Representations 2025},      year={2025},      archivePrefix={arXiv},      eprint={2501.09009},}