Accelerating the Discovery of High-Conductivity Glass Electrolytes via Machine Learning

Glass materials are promising candidates as solid electrolytes for batteries. However, the atomistic origins of variations in ionic conductivity with composition remain poorly understood. A key challenge in computationally designing high-performance glass electrolytes is the absence of an efficient general interatomic potential, particularly for mixed-anion systems, such as combinations of oxides, sulphides, and halides. To address these limitations, a machine learning interatomic potential for glass electrolytes is introduced, based on the atomic cluster expansion descriptor, covering common network formers, modifiers, and anions. Excellent agreement with experimental structure and property data across a wide range of glass compositions is observed. While the potential is broadly applicable to, e.g., lithium-ion conductivity, focus is here on the influence of glass composition on sodium-ion conductivity given the emerging importance of these conductors. By screening the diffusivity of 1,303 glass compositions, the presence of anion species, such as sulphides and halides, is found to have a large positive influence on diffusivity, enabling prediction of highly conductive glasses with mixed anions. Diffusivity and ionic conductivity are highly correlated with a set of descriptors, such as the average atomic volume. This work thus helps to accelerate predictive modelling and broader understanding of conductivity in glass electrolytes.