Bereiche | Tage | Auswahl | Suche | Downloads | Hilfe
SYMS: Symposium Modern developments in multiphysics materials simulations
SYMS 3: Modern developments in multiphysics materials simulations III
SYMS 3.1: Vortrag
Freitag, 29. Februar 2008, 10:15–10:30, A 053
A Neural Network Representation of High-Dimensional Potential-Energy Surfaces for Solids — •Jörg Behler1, Davide Donadio2, Roman Martonak3, and Michele Parrinello4 — 1Lehrstuhl für Theoretische Chemie, Ruhr-Universität Bochum, Germany — 2Department of Chemistry, UC Davis, USA — 3Department of Experimental Physics, Comenius University, Bratislava, Slovakia — 4Department of Chemistry and Applied Biosciences, ETH Zurich, USI-Campus Lugano, Switzerland
We present a generalized neural network scheme for the representation of high-dimensional density-functional theory (DFT) potential-energy surfaces of bulk materials. The neural network is several orders of magnitude faster to evaluate than the underlying DFT energies, while the accuracy of the potential-energy surface is essentially maintained. The method enables long-term molecular dynamics and metadynamics simulations of large systems and is applicable to crystalline and amorphous structures, as well as to the melt. The method is illustrated by studies on the pressure-induced phase transitions in silicon. The obtained results are in excellent agreement with experiment and density-functional theory.