Neural Network Potential Surfaces: A Comparison of two Approaches

Abstract

Abstract Molecular dynamics simulations often involve excessive degrees of freedom, and therefore the computational cost is prohibitively high. Coarse-graining high-dimensional molecular systems is a well-established procedure to increase the space- and time- scales accessible. The coarse-level potential energy must be estimated to derive correctly structural and dynamical properties of the reduced system. Here we examine the artificial neural networks as a potential tool to represent high-dimensional potential-energy surfaces. The simulated coarse configurations are fed to the neural network while taking care to represent each pseudo-atom’s chemical environment properly. Then, the neural network is trained to reproduce the molecular energy. In this work, we present two approaches designed to preserve the physical symmetries of the coarse system. The first approach introduces symmetry functions and the second local coordinates. We implement these techniques, and we apply and compare them for two molecular systems, a bulk methane system, and a liquid water system.