Regensburg 2022 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 10: Poster Session 1
MM 10.3: Poster
Monday, September 5, 2022, 18:00–20:00, P2
Large-scale molecular dynamics simulations using fourth generation neural network potentials — •Emir Kocer1, Andreas Singraber2, Tsz Wai Ko1, Jonas Finkler3, Philipp Misof2, Christoph Dellago2, and Jörg Behler1 — 1Georg-August University, Göttingen, Germany — 2University of Wien, Vienna, Austria — 3University of Basel, Basel, Switzerland
In the last decade, their proven success in bridging the gap between ab initio and classical molecular dynamics made machine learning potentials (MLP) very attractive. However, many MLPs are short-ranged and unable to capture interactions beyond a certain cutoff, which leads to inaccurate forces and energies in systems where long-range interactions are important. While MLPs including long-range electrostatic interactions based on local charges have been available for some years, only recently fourth generation MLPs have 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 step. This study presents a modified version of 4G-HDNNPs, in which the matrix solution is replaced by a function minimization algorithm for an enhanced scalability on multi-core systems. The new potential has been implemented into the LAMMPS software and tested in large-scale molecular dynamics simulations.