Berlin 2018 – scientific programme

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

MM: Fachverband Metall- und Materialphysik

MM 62: Methods in Computational Materials Modelling (methodological aspects, numerics)

MM 62.3: Talk

Thursday, March 15, 2018, 16:15–16:30, TC 006

Improving the training of force-fields based on neural networks — •Mário Marques and Miguel Marques — Institut für Physik, Martin-Luther-Universität Halle-Wittenberg, 06099 Halle (Saale), Germany

The evaluation of potential energy surfaces lies at the heart of many problems in materials science. Density functional theory is often used for this task, but it quickly becomes impractical for systems with hundreds or thousands of atoms.

In this contribution we describe a methodology, based on Behler-Parrinelo approach for artificial neural networks, to solve this problem. We obtain training and test sets from a fully unbiased approach based on global structural prediction techniques.

We extend the back propagation method to consider the error in the forces and stress and we discuss the suitability of a few activation functions. Finally, we develop force fields for Si and Ge, and present some applications to the calculation of point defects and phase diagrams.