Abstract

Iman Mohagheghian : Impact response of polymer-metal laminates Wojciech Szlachta : Gaussian approximation potential for tungsten wjs31@cam.ac.uk

Abstract:

We construct an interatomic potential for tungsten by applying Gaussian process regression to augment an underlying Finnis-Sinclair potential with the missing quantum-mechanical (QM) information. The QM information is computed directly using density functional theory (DFT), and the difference between the potential energy surfaces is interpolated by the Gaussian process. As a result, we obtain an interatomic potential which is no longer constrained by a fixed functional form, and is therefore capable of representing the complex potential energy landscape with high accuracy.

We represent atomic environments using a set of rotationally and permutationally invariant parameters, which act as independent variables in our regression equations. Our training data consists of the most common crystal phases over a range of volumes and strains, as well as simple lattice defects, such as stacking faults / gamma surfaces, vacancies and surfaces.

We find that the properties of our new potential are in agreement with the DFT data, and that the force and energy errors of the underlying Finnis-Sinclair potential are significantly reduced. Our new potential can also reproduce the structure of the screw dislocation correctly, even though it was not included in our training data set.