Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 62: Developement of Calculation Methods III

MM 62.6: Talk

Thursday, March 21, 2024, 17:00–17:15, C 264

Cross-Platform Hyperparameter Optimizer for Machine-Learning Potential Fitting — •Daniel F. Thomas du Toit, Yuxing Zhou, and Volker L. Deringer — Department of Chemistry, University of Oxford, Oxford, UK

The use of machine learning interatomic potentials (MLIPs) to study materials has become increasingly popular in recent years. As the field has matured, multiple frameworks for MLIP fitting have been proposed. Here we present a Python package to optimize hyperparameters for MLIPs.

Our package, XPOT (``Cross-platform optimizer for machine learning interatomic potentials"), uses bespoke interfaces to MLIP fitting programs. XPOT enables users to use automated optimization to fit robust, accurate, and fast MLIPs. Using XPOT, we optimized hyperparameters for SNAP and ACE potentials based on existing training databases for Gaussian approximation potential (GAP) models, and demonstrated cost improvements while retaining high accuracy. We showcase the usefulness of the approach by creating optimized MLIPs across a diverse range of complex materials systems.

Keywords: machine learning; interatomic potentials; phase-change materials; molecular dynamics; software