Regensburg 2013 – wissenschaftliches Programm

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

O: Fachverband Oberflächenphysik

O 58: Poster Session III (Solid-liquid interfaces; Scanning probe and other methods; Electronic structure theory; Spin-orbit interaction)

O 58.75: Poster

Mittwoch, 13. März 2013, 18:15–21:45, Poster B1

Representing Complex Potential Energy Surfaces by Artificial Neural Networks — •Christopher Handley, Tobias Morawietz, and Jörg Behler — Lehrstuhl für Theoretische Chemie, Ruhr-Universität Bochum, D-44780 Bochum, Germany

Simulations of large systems using ab initio methods are computationally costly, and in many cases intractable. An alternative is approximate interatomic potentials. These potentials are typically constructed from computationally simple functions, chosen to be representative of particular atomic interactions. Fitting of these potentials is not straightforward, as many different selections of functions and their parameters are valid, with varying accuracy. Neural Networks (NNs) recently have been shown to provide interatomic potentials that are comparable to the accuracy of quantum mechanical calculations[1,2]. NNs are flexible enough to fit complex functions, interpolating, from quantum mechanical training data, accurate energies and forces. A major drawback of the method is the non-physical functional form of the NNs. Here, we present the first steps towards a more transferable NN based upon electronic structure methods.

[1] C. M. Handley and P. L. A. Poplier, J. Phys. Chem. A, 114, 3371-3383, (2010).

[2] J. Behler, PCCP, 13, 17901-18232 (2011).