Dresden 2017 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 60: Topical session: Data driven materials design - machine learning
MM 60.2: Vortrag
Donnerstag, 23. März 2017, 12:15–12:30, BAR 205
Representing energy landscapes by combining neural networks and the empirical valence bond method — •Sinja Klees1, Ramona Ufer2, Volodymyr Sergiievskyi2, Eckhard Spohr2, and Jörg Behler1 — 1Lehrstuhl für Theoretische Chemie, Ruhr-Universität Bochum, D-44780 Bochum, Germany — 2Lehrstuhl für Theoretische Chemie, Universität Duisburg-Essen, D-45141 Essen, Germany
In recent years, artificial neural networks (NNs) have become a powerful method to develop reactive interatomic potentials for a wide range of systems. Due to their high flexibility, they allow to interpolate reference energies and forces obtained from electronic structure calculations, without the introduction of any constraint to the functional form. However, the construction of NN potentials can become computationally very demanding due to the high dimensionality of the configuration space, which needs to be mapped. Combining NN potentials with the empirical valence bond (EVB) method offers a promising approach to derive the potential energy surface of complex systems with substantially reduced effort, since the size of the reference structures can be strongly decreased by employing the EVB method to combine smaller fragments in a physically meaningful way. Preliminary results will be discussed and compared to density functional theory data.