SurfaceScience21 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 42: Poster Session III: Poster to Mini-Symposium: Machine learning applications in surface science I
O 42.2: Poster
Dienstag, 2. März 2021, 10:30–12:30, P
A fourth-generation high-dimensional neural network potential — •tsz wai ko1, jonas a. finkler2, stefan goedecker2, and jörg behler1 — 1Theoretische Chemie, Institut für Physikalische Chemie, Georg-August-Universität Göttingen, Tammannstr. 6, 37077 Göttingen, Germany — 2Department of Physics, Universität Basel, Klingelbergstrasse 82, 4056 Basel, Switzerland
Machine learning potentials (MLPs) have become an important tool for performing reliable atomistic simulations in surface science due to their nearly ab-initio accuracy and efficiency comparable to empirical force field. The majority of MLPs relies on the representation of energies and sometimes charges as a function of the local atomic environments. They are thus unable to describe non-local changes in the electronic structure due to long-range charge transfer or different global charges of a system.
Here we proposed a fourth-generation high-dimensional neural network potential (4G-HDNNP) for capturing the global charge distributions and corresponding non-local effects. We demonstrate the performance of 4G-HDNNPs for different benchmark systems showing that 4G-HDNNPs are in excellent agreement with electronic structure calculations.