Deep learning potential for superionic phase of Ag2S

Abstract

Artificial neural networks are used for describing potential energy surface of β-Ag2S silver sulfide. It has allowed performing accurate and fast atomistic simulations for describing behavior of investigated system. We develop neural network potential for high temperature ionic conductor β-Ag2S using DeePMD approach. Reference ab initio dataset was generated using active learning technique implemented in DP-GEN package. Classical molecular simulations with developed neural network potential were performed. Partial radial distribution function for S-S pair and bond-angle distribution function for S-S-S triplet demonstrate crystalline behavior, while the same functions for Ag-Ag pair and Ag-Ag-Ag triplet demonstrate liquid-like behavior. Mean squared displacement of S atoms indicates absence of diffusion for sulfur atoms, while the same function for Ag atoms has linear form at large times that indicates presence of diffusion for this sort of atoms. Velocity autocorrelation functions for S atoms have oscillatory behavior, while for Ag atoms no oscillations are observed. Comparison of mean squared displacement for S atoms and diffusivity for Ag atoms is performed to other ab initio and classical simulations as well as experimental data and demonstrates good agreement in all the cases. Obtained by active learning technique dataset could be expanded to other Ag2S phases for describing Ag2S in wider range of temperatures. Thus accurate, productive, almost free of parameters and promising for future use model for β-Ag2S was created.