Prediction of potential energy profiles of molecular dynamic simulation by graph convolutional networks

Abstract

A graph convolutional networks (GCN)-based machine learning (ML) model is constructed to predict physical properties of metallic materials from graph representation of atomic configuration of molecular dynamics (MD) simulation. The developed ML model is employed for the prediction of time variation of the potential energy of a solid–liquid biphasic system of nickel. The learned ML model gives a good prediction on the property of training data. Moreover, it is confirmed that the ML model has generalization performance sufficient to make adequate predictions on unknown graph structures despite the lack of information on interatomic distances in the graph representation. It is significant in this study to show that the graph representation can be a good notation for the prediction of various properties from MD simulations since there is no established notation for atomic configuration of MD simulation especially for large-scale system of metallic materials.