Prediction of ionic conductivity in solid-state electrolytes using machine learning

Data and models to predict cation conductivity are crucial for designing solid-state electrolytes for rechargeable batteries, with Li and Na being among the most widely studied cations. To support such studies, we used large language models (LLMs) and manual data curation to construct a large, systematically curated dataset of room-temperature ionic conductivity for Li- and Na-based solid electrolytes, suitable for composition-based machine learning. We then applied machine learning models, such as Random Forest Regressor (RF), eXtreme Gradient Boosting (XGBoost, XGB), Gradient Boosting Regressor (GB), Multi-Layer Perceptron (MLP), and Stacking Regressor to predict the logarithm of conductivity from elemental properties. Among the evaluated models, the Random Forest model exhibited superior performance across all datasets, achieving an R 2 of 0.569/MAE of 0.911 for the Li dataset, an R 2 of 0.691/MAE of 0.665 for the Na dataset, and an R 2 of 0.598/MAE of 0.864 for the combined Li-Na dataset. These results indicated only limited quantitative predictive capability, but the model could effectively differentiate good ( ≥ 10⁻⁴ S/cm) from poor conductors, achieving a maximum F1 score of 0.83 for the Li dataset, 0.92 for the Na dataset, and 0.85 for the combined Li and Na dataset. Using SHapley Additive exPlanations (SHAP), we identified bonding strength and atomic size characteristics as the most important features for predicting ionic conductivity in our model.