Fine-tuning MLIP foundation models: strategies for accuracy and transferability
Paper • 2606.12704 • Published
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('imagefolder', {}), NamedSplit('test'): ('csv', {})}
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
MACE Fine-Tuning Supplementary
Supplementary data and scripts for:
Tompa, T. L.; Varga-Umbrich, E.; Batatia, I.; Elena, A. M.; Bernstein, N.; Csányi, G. Fine-tuning MLIP foundation models: strategies for accuracy and transferability (2026). arXiv:2606.12704.
The repository contains training datasets, mace_run_train launch scripts, Slurm logs, fine-tuned model checkpoints (.model), evaluation scripts, and processed results for the paper. Benchmark systems: lithium argyrodite electrolytes (li), aqueous NaCl (nacl), ice polymorphs (water), SN₂ reactions (sn2), and MACE-OFF biomolecules (spice). The E0s_MH1_model/ tree holds the MACE-MH1 isolated-atom E0 reestimation study on UiO-66.
Repository layout
MACE_finetuning_supplementary/
├── foundation-utils/ # Foundation checkpoints and OMat24 replay pools
├── E0s_MH1_model/ # MACE-MH1 E0 reestimation on UiO-66
├── li/ # Lithium electrolyte (LPSC)
├── nacl/ # Aqueous NaCl
├── water/ # Ice polymorph
├── sn2/ # SN₂ reaction
├── spice/ # SPICE biomolecules
├── rss-eval/ # Random structure search (short-range repulsion) evals
└── ablations/ # Hyperparameter and design ablations
Licensing
Scripts, logs, checkpoints, and results produced for this study are under LICENSE (CC BY 4.0, Tompa et al.). Dataset structure files retain upstream licenses in THIRD_PARTY_LICENSES.md.
| Data path | License | Source |
|---|---|---|
li/data/ |
CC BY 4.0 | Zenodo 15686940 (Kim et al., reEWC) |
spice/data/ |
MIT | 10.17863/CAM.107498 (MACE-OFF) |
nacl/data/ |
CC BY-SA 4.0 | niamhon/nacl-water |
water/data/ |
CC BY 4.0 | fine-tuning-MLPs-ice-polymorphs |
sn2/data/ |
CC BY 4.0 | organic_neb_mace |
foundation-utils/omat24_filtered_*.extxyz |
CC BY 4.0 | OMat24 (subsamples) |
li/data/mp_traj*.xyz |
CC BY 4.0 | Materials Project / MPtrj (subsamples) |
E0s_MH1_model/data/*uio66* |
CC BY 4.0 | Elena et al., mace-mof-0 (UiO-66 structures) |
E0s_MH1_model/data/omat*-replay* |
CC BY 4.0 | OMat24 (replay subsamples) |
Citation
@misc{tompa2026mlipfinetuning,
title = {Fine-tuning {MLIP} foundation models: strategies for accuracy and transferability},
author = {Tompa, Tam{\'a}s Lajos and Varga-Umbrich, Eszter and Batatia, Ilyes and Elena, Alin M. and Bernstein, Noam and Cs{\'a}nyi, G{\'a}bor},
year = {2026},
eprint = {2606.12704},
archivePrefix = {arXiv},
primaryClass = {physics.chem-ph},
url = {https://arxiv.org/abs/2606.12704},
note = {Journal: TBD}
}
See THIRD_PARTY_LICENSES.md for upstream data citations.
Authors
Tamás Lajos Tompa, Eszter Varga-Umbrich, Ilyes Batatia, Alin M. Elena, Noam Bernstein, Gábor Csányi
- Downloads last month
- 1,323