Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials

Supplementary data for the paper Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials by Danny Perez, Aparna P. A. Subramanyam, Ivan Maliyov, and Thomas D. Swinburne [arxiv 2025](https://arxiv.org/abs/2502.07104) @@article{perez2025uncertainty, title={Uncertainty Quantification for Misspecified Machine Learned Interatomic Potentials}, author={Perez, Danny and Subramanyam, Aparna and Maliyov, Ivan and Swinburne, Thomas D}, journal={arXiv preprint arXiv:2502.07104}, year={2025} } Contents of the repository RAW_DATA: Training data from the entropy-maximization dataset, stored in FitSNAP json format. FEATURIZED_DATA: Featurization of the raw data in terms of bispectrum components used in qSNAP. Obtained with the FitSNAP code. RESAMPLED_POTENTIALS: 500 SNAP potentials resampled from the hyper-cube bounding the POPS ensemble, stored in LAMMPS-readable snapparams/snapcoeff format. Also includes the full POPS ensemble. RESAMPLED_POTENTIALS_CHARACTERIZATION: Characterization of the resampled potentials on the materials properties presented in the manuscript. Note that these currently cannot be reproduced externally due to the use of non-public analysis libraries. These are provided for reference only. NOTEBOOKS: NOTEBOOKS/fit-models.ipynb Notebook used to process the featurized data, fit the qSNAP models, generate the POPS ensemble, and resample from the hyper-cube bounding the POPS ensemble. NOTEBOOKS/POPS_MACE_TEST Notebook illustrating demonstration of the POPS approach to the MACE universal potential.