Strategy for building training set for neural-network potentials for ionic diffusion in solids: example for hydride-ion diffusion in LaHO

Solid ionic conductors represent an important class of materials for many practical applications, such as fuel cells, batteries and hydrogen storage. Molecular dynamics (MD) modeling is a powerful tool to perform large-scale simulations to study ionic trajectories at the atomic level. However, since MD is critically dependent on the potential of interatomic interactions, it is of utter importance to have the potential that can successfully reproduce experimental aspects of ionic diffusion, such as conductivity and activation energy. In this paper, we present our strategy for building a training set for neural-network potentials (NNPs) that allow to attain a good agreement between MD and experimental results. To prove validity of the strategy, we developed NNP for LaHO, novel H- conductor and explicitly illustrated that the diffusion of the H- ions in LaHO is attributed to the defect formation and migration. [GRAPHICS]