Molecular dynamics simulations of lithium superionic conductor Li10GeP2S12 using a machine learning potential

The molecular dynamics simulations have been performed for tetragonal Li10GeP2S12 (LGPS) to investigate the Li-ion conductivity and underlying diffusion mechanism. The atomic interactions are simulated by the machine learning potential which provides accuracy comparable to ab-initio calculations and enable to collect statistics at root temperature. The obtained Li sites are consistent with the four-site model identified by the experiments. The Li-ion conductivity at 300 K and the activation energy are predicted to be σ = 12 mS/cm and Ea = 226 meV, respectively, which are in remarkable agreement with the experimental data. It is found that the ratio of the Li-ion diffusivity in the c direction to that in the ab plane is approximately three and the activation energies for both components are almost identical, suggesting a small anisotropy for Li-ion diffusion in LGPS. Our numerical experiment clearly indicates that the diffusion along the in-plane two-dimensional paths is the rate-limiting process in the overall diffusion.