Regensburg 2010 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Downloads | Hilfe

O: Fachverband Oberflächenphysik

O 60: Poster Session III (Methods: Atomic and electronic structure; Methods: electronic structure theory; Methods: Molecular simulations and statistical mechanics; Methods: Sanning probe techniques; Methods: other (experimental); Methods: other (theory) )

O 60.5: Poster

Mittwoch, 24. März 2010, 17:45–20:30, Poster B2

Analysis of Neural Network Potential-Energy Surfaces for Atomistic Simulations — •Tobias Morawietz, Nongnuch Artrith, and Jörg Behler — Lehrstuhl für Theoretische Chemie, Ruhr-Universität Bochum, D-44780 Bochum, Germany

In recent years, artificial Neural Networks (NN) have become a promising new method to represent accurate potential energy surfaces (PES) for molecular and condensed systems. NN potentials are constructed using a set of reference points obtained in electronic structure calculations, which are then interpolated to a continuous PES by the very flexible functional form of the NNs. The NN potentials can then be evaluated several orders of magnitude faster than the underlying electronic structure data. Therefore NN potentials are ideal tools to perform long molecular dynamics simulations of large systems. However, because of the a priori non-physical functional form, NN potentials have to be constructed with care. Using various molecules, clusters and solids we analyze how the physical functional form is approximated by NNs and discuss the scope and limitations of the method.