Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 4: Data Driven Material Science: Big Data and Workflows I
MM 4.1: Vortrag
Montag, 18. März 2024, 10:15–10:30, C 243
Investigating Structural Descriptors for High-Dimensional Neural Network Potentials — •Moritz R. Schäfer1,2, Moritz Gubler3, Stefan Goedecker3, and Jörg Behler1,2 — 1Theoretische Chemie II, Ruhr-Universität Bochum, Germany — 2Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr, Germany — 3Department of Physics, University of Basel, Switzerland
High-dimensional neural network potentials (HDNNPs) are a well established technique to efficiently compute energies and forces akin to ab initio standards for conducting extensive molecular dynamics simulations of intricate systems in high dimensions. This method expresses the total energy from environment-specific atomic energy contributions, with the option to incorporate electrostatic interactions utilizing flexible atomic charges. The reliability of both components significantly depends on the accuracy of the structural descriptors used to define the atomic environments. Here, we combine atom-centered symmetry functions with the newly introduced overlap matrix descriptor. Furthermore, we analyze the strengths and weaknesses of each descriptor, providing insights through demonstrations on benchmark systems.
Keywords: Machine Learning Potentials; HDNNPs; Descriptors; Simulation; MD