Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 12: Poster I
MM 12.18: Poster
Monday, March 27, 2023, 18:15–20:00, P2/OG1+2
How to Train a Neural Network Potential — •Alea Miako Tokita1,2 and Jörg Behler1,2 — 1Lehrstuhl für Theoretische Chemie II, Ruhr-Universität Bochum, 44780 Bochum, Germany — 2Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr, 44780 Bochum, Germany
High-Dimensional Neural Network Potentials (HDNNPs) provide potential energy surfaces (PESs) with the accuracy of electronic structure calculations at strongly reduced computational costs, which enables extended molecular dynamics simulations of large systems. They are trained on reference energy and force data to learn an approximate but accurate functional relation between the atomic structure and the PES. However, due to the non-physical functional form, which is shared with many other types of machine learning potentials, this training and the validation of the potential have to be done with great care. In this contribution the construction of HDNNPs will be explained step by step including a discussion of possible pitfalls and tricks of the trade.