Dresden 2011 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

O: Fachverband Oberflächenphysik

O 22: Theoretical methods

O 22.2: Talk

Monday, March 14, 2011, 17:30–17:45, WIL C107

Representing Potential-Energy Landscapes by High-Dimensional Neural Networks — Nongnuch Artrith, Tobias Morawietz, and •Jörg Behler — Lehrstuhl für Theoretische Chemie, Ruhr-Universität Bochum, 44780 Bochum, Germany

The construction of efficient interatomic potentials is a necessary condition for carrying out large scale molecular dynamics and Monte Carlo simulations. However, the identification of a suitable functional form to represent a given set of reference data from ab initio calculations with high accuracy is a formidable challenge.

Using several benchmark systems, we show that flexible artificial neural networks (NNs) are able to provide reliable high-dimensional potential-energy surfaces even if complex bonding situations are present. They can be applied to a wide range of systems from bulk metals and surfaces to small molecules. The scope and limitations of NN potentials are discussed, and the accuracy is compared with electronic structure calculations.