SKM 2023 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

O: Fachverband Oberflächenphysik

O 26: Poster: New Methods

O 26.6: Poster

Montag, 27. März 2023, 18:00–20:00, P2/EG

RuNNer 2.0: An efficient and modular program for training and evaluating high-dimensional neural network potentials — •Alexander L. M. Knoll^{1, 2}, Marco Eckhoff³, Jonas A. Finkler⁴, Tsz Wai Ko⁵, Emir Kocer^{1, 2}, K. Nikolas Lausch^{1, 2}, Moritz R. Schäfer^{1, 2}, Gunnar Schmitz^{1, 2}, Alea Miako Tokita^{1, 2}, and Jörg Behler^{1, 2} — ¹Lehrstuhl für Theoretische Chemie II, Ruhr-Universität Bochum, 44780 Bochum, Germany — ²Research Center Chemical Sciences and Sustainability, Research Alliance Ruhr, 44780 Bochum, Germany — ³ETH Zürich, Laboratorium für Physikalische Chemie, 8093 Zürich, Switzerland — ⁴Department of Physics, Universität Basel, 4056 Basel, Switzerland — ⁵Department of NanoEngineering, University of California, San Diego, CA, USA

Machine learning potentials (MLPs) have become a popular tool for large-scale atomistic simulations in chemistry and materials science. They provide efficient access to highly accurate potential energy surfaces (PES) generated from ab initio reference calculations. As methods in this field are becoming increasingly complex and reach maturity, the development of efficient and user-friendly tools is also gaining importance. We present the second major release version of RuNNer, an open source, stand alone software package for the construction and evaluation of second, third, and fourth generation high-dimensional neural network potentials (HDNNPs). RuNNer 2.0 unifies an entire workflow in a fully MPI-parallel program: from the generation of atomistic descriptors, over the training of a specific machine learning model, to its final application in molecular dynamics codes.