Regensburg 2016 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 55: Methods in Computational Materials Modelling III: Machine learning and statistics
MM 55.1: Talk
Thursday, March 10, 2016, 11:45–12:00, H53
Gaussian approximation potentials: the case of α-iron — •Daniele Dragoni1, Tom Daff2, Gábor Csányi2, and Nicola Marzari1 — 1Theory and Simulation of Materials (THEOS), and National Center for Computational Design and Discovery of Novel Materials (MARVEL), École polytechnique fédérale de Lausanne, 1015 Lausanne, Switzerland — 2Engineering Laboratory, University of Cambridge, Trumpington Street, Cambridge, CB2 1PZ, United Kingdom
Interatomic potentials are typically based on functional forms driven by physical intuition and fitted to experimental or computational data. The moderate flexibility of these functional forms limits their ability to be systematically improved by increasing the fitting datasets, although ensuring a modicum degree of transferability. Recently, a novel trend has emerged where potential-energy surfaces are represented by neural networks fitted on large numbers of first-principles calculations, thus maximizing flexibility but requiring extensive datasets to ensure transferability. Gaussian Approximation Potentials in particular are a novel class of potentials based on non-linear, non-parametric Gaussian-process regression. We apply this approach to the case of α-iron, training a GAP model from energies, stresses and forces taken from first-principles molecular dynamics simulations of pristine and defected bulk systems, and of surfaces with different crystallographic orientations, covering roughly 105 local atomic environments. Finally, we verify this GAP model by comparing its predicted structural, vibrational, and thermodynamic properties against those derived from first-principles.