Neural Network-Based Simulation Method to Examine Ion Behaviors under Electric Fields: Application to Ion Migration in Amorphous Li₃PO₄

Abstract

We developed a neural network-based model to predict the Born effective charges from atomic structures. By combining forces due to an applied electric field, expressed as a product of the Born effective charge and the electric field, and forces evaluated by a neural network potential (NNP), a simulation scheme of ion dynamics under an electric field was proposed. Taking Li3PO4 as a prototype, we demonstrated the validity of our computation scheme. Using the constructed model of the Born effective charge predictor and NNP based on density functional (perturbation) theory calculation data, molecular dynamics (MD) simulations under a uniform electric field of 0.1 V/Å were performed. We obtained an enhanced mean square displacement of Li along the electric field, which seems physically reasonable. In addition, we found that the external forces along the direction perpendicular to the electric field, which originated from the off-diagonal components of the Born effective charges, had a non-negligible effect on the Li motion. Furthermore, we observed a more susceptive response of Li to the electric field in an amorphous structure.