Prediction of the low-temperature properties of electrolyte solvents for lithium-ion batteries via machine learning

Abstract

Electrolytes with low melting points (MPs), high boiling points (BPs), and high dielectric constants (ε) can effectively mitigate performance degradation in lithium-ion batteries (LIBs) under low-temperature conditions. However, the lack of systematic experimental data on electrolyte properties poses a significant challenge to traditional design approaches. To address this limitation, we developed a machine learning workflow that integrates data acquisition using large language models, model construction, and interpretability analysis, aiming to predict key molecular properties, with a focus on MPs, BPs and ε. We constructed a multi-source database, LiElectroDB, that contains over 150 000 electrolyte molecules relevant to LIBs. The prediction models demonstrate strong performance across all three properties, achieving an R² of 0.8864 and a root mean square error (RMSE) of 23.3 K for the MP, a coefficient of determination (R²) of 0.9608 and an RMSE of 14.3 K for the BP using the XGBoost algorithm, and an R² of 0.8718 and a RMSE of 6.7 for ε using an artificial neural network. To further uncover structure–property relationships, t-SNE and SHAP are employed to analyze the molecular features contributing to thermal behavior at a microscopic level. Finally, by integrating molecular neighborhood search with high-throughput screening, nine candidate molecules are identified as promising low-temperature electrolytes for LIBs. This work provides an efficient and generalizable framework for the design of low-temperature electrolytes in LIBs.