SKM 2023 – scientific programme

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

MM: Fachverband Metall- und Materialphysik

MM 25: Development of Computational Methods: Crystal Structure and Properties

MM 25.4: Talk

Wednesday, March 29, 2023, 11:00–11:15, SCH A 251

Efficient molecular dynamics simulations using fourth-generation neural network potentials — •Emir Kocer^1,2, Andreas Singraber³, Tsz Wai Ko^1,2, Jonas Finkler⁴, Philipp Misof³, Christoph Dellago³, and Jörg Behler^1,2 — ¹Lehrstuhl für Theoretische Chemie II, Ruhr-Universität Bochum, Bochum, Germany — ²Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr, Bochum, Germany — ³Institut für Experimentalphysik, Universität Wien, Wien, Austria — ⁴Departement Physik, Universität Basel, Basel, Switzerland

Machine learning potentials (MLP) have become a common tool in materials modelling due to their ability to bridge the gap between ab initio and classical molecular dynamics. A limitation of most MLPs, however, is the locality approximation in that only interactions within a cutoff range are considered. This could lead to inaccurate dynamics in systems with relevant long-range interactions. Recently, a fourth-generation of MLPs has emerged that can take also global phenomena like non-local charge transfer into account. An example is the fourth-generation high-dimensional neural network potential (4G-HDNNP), which utilizes a global charge equilibration. In this study, a modified version of 4G-HDNNPs with enhanced efficiency will be presented. The new algorithm has been implemented in the LAMMPS software and tested in large-scale molecular dynamics simulations.