Machine Learning-Aided Materials Design Platform for Predicting the Mechanical Properties of Na-Ion Solid-State Electrolytes

Abstract

Na-ion solid-state electrolytes (Na-SSEs) exhibit high potential for electrical energy storage owing to their high energy densities and low manufacturing cost. However, their mechanical properties that are critical to maintaining structural stability at the interface are still insufficiently understood. In this study, a machine learning-based regression model was developed for predicting the mechanical properties of Na-SSEs. As a training set, 12,361 materials were obtained from a well-known materials database (Materials Project) and were represented with their respective chemical and structural descriptors. The developed surrogate model exhibited remarkable accuracies (R-2 score) of 0.72 and 0.87, with mean absolute errors of 11.8 and 15.3 GPa for the shear and bulk modulus, respectively. This model was then applied to predict the mechanical properties of 2432 Na-SSEs, which have been validated with first-principles calculations. Finally, the optimization process was performed to develop an ideal materials screening platform by adding the minimized data set, wherein the prediction uncertainty is reduced. We believe that the platform proposed in this study can accelerate the search for Na-SSEs with ideal mechanical properties at minimum cost.