Evaluating Descriptor Influence on Training Data Selection and Performance in Machine-Learned Interatomic Potentials

Machine-learned interatomic potentials (MLIPs) have become essential for materials modeling, providing a capability for atomistic simulations with near ab initio accuracy at a fraction of the cost. These models rely critically on the quality and completeness of their training data. Past efforts have focused on quantifying these dataset properties through “general purpose” local environment descriptors. Since different descriptors produce dataset latent spaces that vary even for the same underlying dataset, we hypothesize that dataset quality can be improved by taking into consideration the specific model being trained, enabling construction of smaller yet more information rich MLIP training sets.

We present efforts to probe this hypothesis in terms of how these descriptor-defined latent spaces influence MLIP training data construction and model performance. Using a benchmark dataset developed for cross-model comparison [1], we prune configurations in latent spaces constructed using descriptors of the ACE, GAP, NNP, ChIMES, and SNAP MLIPs. Each model is then retrained on (i) datasets pruned using its own descriptor and (ii) data pruned using latent spaces from other descriptors.

Resulting models are evaluated on the basis of test errors, physical properties, and stable dynamics. We assess how descriptor-dependent data pruning impacts predictive performance. This study provides a direct comparison of descriptor efficacy for data selection and highlights their role in shaping the balance between model accuracy and transferability.

This work was performed in part under the auspices of the U.S. Department of Energy by Lawrence Livermore National. Laboratory under Contract DE-AC52-07NA27344.