Regensburg 2025 – scientific programme

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

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 15: Poster Session I

CPP 15.65: Poster

Monday, March 17, 2025, 19:00–21:00, P4

Insights into Machine Learning Interatomic Potentials for simple analytical model systems — •Mirko Fischer and Andreas Heuer — Institute for Physical Chemistry, University of Münster, Corrensstraße 28/30, 48149 Münster

During the past 10 years Machine Learning interatomic potentials (MLIP) have gained popularity for Molecular Dynamics simulations with quantum chemical accuracy. Although it is a rapidly evolving field, many questions remain open. These include issues of interpretability, modeling different interaction types and how to select training data properly. MLIPs are rarely applied to simple model systems, for which the interactions of particles can be described analytically, to investigate these questions. Instead, MLIPs are mostly applied directly to realistic molecular systems for which the ground truth interactions must be approximated by methods like Density Functional Theory. By training an Atomic Cluster Expansion as a systematic and generally interpretable model for a Lennard-Jones model system, we aim to study how interaction types, system size and temperature affect the learned model in a systematic manner. The obtained interactions can be directly compared to the known true analytical interactions. Moreover, we fit a MLIP for an amorphous silica system to study structure and dynamics in a glass-forming system, where low-energy states are important. The question emerges, how to select the training data best and if such low-energy (low temperature) states must be explicitly included in it or if the MLIP is able to extrapolate from high-energy (high temperature) states to low energy states.

Keywords: Atomic Cluster Expansion; Machine Learning; Molecular Dynamics