Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 58: Poster Wednesday: New Methods and Developments, Frontiers of Electronic Structure Theory
O 58.5: Poster
Mittwoch, 7. September 2022, 18:00–20:00, P4
Assessment of Structural Descriptors for the Construction of High-Dimensional Neural Network Potentials — •Moritz R. Schäfer1, Jonas Finkler2, Stefan Goedecker2, and Jörg Behler1 — 1Georg-August-Universität Göttingen, Institut für Physikalische Chemie, Theoretische Chemie, Tammannstraße 6, 37077 Göttingen, Germany — 2Basel University, Department of Physics and Astronomy, Klingelbergstrasse 82, 4056 Basel, Switzerland
High-dimensional neural network potentials (HDNNPs) can be used to efficiently compute close-to ab initio quality energies and forces for performing large-scale molecular dynamics simulations of complex systems. In this method, the total energy is constructed as a sum of environment-dependent atomic energy contributions. Also electrostatic interactions based on environment-dependent charges can be included. Hence, a set of reliable structural descriptors for the atomic local environments is crucial to develop accurate potentials. Often, atom-centered symmetry functions (ACSFs) are used for this purpose in HDNNPs. In this work, we benchmark the accuracy and transferability of HDNNPs with respect to alternative descriptors like the recently proposed overlap matrix descriptor.