Physics-informed machine learning for Na-Ion conductivity and activation energy

Abstract

Glass-based electrolytes are promising for solid-state batteries due to the absence of grain boundaries. However, filtering the compositional space for suitable glass formulations is challenging due to the extremely wide compositional space of multicomponent glasses. Here, we consider a highly consistent dataset of sodium-ion conductive glasses for developing a physics-informed machine learning model (PIML) of ion conductivity and transport activation energy. The PIML achieves a remarkable 99% goodness of fit for conductivity, outperforming classical ML models. Moreover, it predicts conductivity $\left(\sigma \right)$, activation energy ($E_a)$, and pre-exponential factor $\left({\sigma }_0\right)$ at various temperatures, even for compositions beyond the training data. Incorporating the Arrhenius equation as a hard constraint enables physically meaningful predictions, especially at varying temperatures. Interpreting the PIML model using Shapley additive explanations (SHAP) highlights oxygen's substantial impact on $E_a$. Aside from providing progress in the specific field of glass-based solid electrolytes, these findings highlight the ongoing importance of consistent experimental campaigns for high-quality ML predictions.