Information-entropy-driven generation of material-agnostic datasets for machine-learning interatomic potentials

• DOI: 10.48550/arxiv.2407.10361
• Publication Date: 2024-07-14
• Authors: Aparna P. A. Subr...; Danny Pérez
• Keywords: Condensed Matter - Materials Science; Materials Science (cond-mat.mtrl-sci); FOS: Physical sciences

In contrast to their empirical counterparts, machine-learning interatomic potentials (MLIAPs) promise deliver near-quantum accuracy over broad regions of configuration space. However, due generic functional forms and extreme flexibility, they can catastrophically fail capture the properties novel, out-of-sample configurations, making quality training set a determining factor, especially when investigating materials under conditions. present study, we propose novel automated dataset generation method based on maximization information entropy feature distribution, aiming at an extremely coverage space in way that is agnostic specific target materials. The ability unique material demonstrated range unary materials, including elements with fcc (Al), bcc (W), hcp (Be, Re Os), graphite (C) ground states. MLIAPs trained this are shown be accurate application-relevant metrics, as well robust very swaths configurations space, even without fine-tuning or hyper-parameter optimization, approach attractive rapidly autonomously develop general-purpose suitable for simulations