Study of diffusion and conduction in lithium garnet oxides LixLa3Zrx−5Ta7−xO12 by machine learning interatomic potentials

Abstract

Lithium garnet oxides are an attractive family of solid-state electrolytes due to their high Li-ion conductivity and good chemical stability against Li metal. Experimental study of these materials is often troubled by chemical contamination (e.g. Al) or lithium loss, while computational study, theoretically with controlled composition, is often limited either by accuracy (e.g. conventional interatomic potential) or efficiency (e.g. density-function theory or DFT). In this work, we report the study of diffusion and conduction of lithium garnets by a machine learning interatomic potential (MLIP) that is approaching DFT accuracy but orders of magnitude faster. We found that this MLIP is more accurate than other commonly applied models to study lithium garnets and is able to predict diffusion and conduction properties that are consistent with experiments. Computational studies enabled by this MLIP, combined with experimental data, suggest that ionic conduction is non-Arrhenius and maximum conductivity occurs around x = 6.6 to 6.8 in Li_xLa₃Zr_x−5Ta_7−xO₁₂.