Machine Learning-Assisted Design of Na-Ion-Conducting Glasses

Abstract

As an alternative to liquid electrolytes, all-solid-state sodium-ion batteries are receiving significant attention due to their potential for improved safety and efficiency. Here, we propose a combined experimental and machine learning (ML) approach for discovering glass electrolytes while also providing insights into the role of different glass components. Specifically, we experimentally prepare and measure the ionic conductivity of 27 glass compositions of the sodium aluminophosphate glass family. Further, we train ML models on this dataset to predict the ionic conductivity, which exhibits excellent agreement with the experimental results. We interpret the composition–conductivity relationship learned by the ML model using Shapely additive explanations (SHAP), which reveals the role played by the glass components in governing the conductivity. Employing these observations, glass compositions with improved conductivity values are predicted and experimentally validated. The results corroborate the insights from SHAP analysis and enable optimized glass formulations in real-world experiments. This demonstrates how ML tools can significantly accelerate the discovery of Na-ion-conducting glass electrolytes.