Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 18: New methods: Theory – Poster
O 18.3: Poster
Montag, 9. März 2026, 18:00–20:00, P2
Increasing the Transferability of Machine Learning Potentials by Learning Atomic Properties — •Johann Richard Springborn1,2, Gunnar Schmitz1,2, and Jörg Behler1,2 — 1Theoretische Chemie II, Ruhr-Universität Bochum, Germany — 2Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr, Germany
Machine Learning Potentials (MLPs) have become an established tool for describing potential energy surfaces of complex systems. While they significantly speed up the evaluation of the energy and forces in comparison to ab-initio methods, they require high-quality reference data for training. Depending on the systems to be studied, generating this training data can become the computational bottleneck. Therefore it is of high interest to reduce the number and complexity of structures to be computed. We propose to achieve this goal by training MLPs on atomic properties instead of global quantities such as the system's total energy. This approach aims to increase the transferability of the resulting MLPs to more complicated systems while still utilizing easily accessible reference data.
Keywords: Machine learning potentials