Interpretable enhanced-ECFP-guided deep learning for rational electrolyte design and Coulombic efficiency prediction in lithium metal batteries

초록

Rational electrolyte design is essential for enhancing the Coulombic efficiency (CE) and interfacial stability of lithium metal batteries. However, current molecular representations limited in their description of molecular connectivity and lacking substructure frequency information constrain the development of accurate and interpretable AI models. This paper introduces an interpretable AI framework to predict CE using an enhanced extended-connectivity fingerprint representation that integrates molecular type, substructure frequency, and concentration weighting. A dataset of 168 CE values from Li-Cu half-cells was used to train a deep neural network, which outperformed conventional methods in predictive accuracy. SHapley Additive exPlanations analysis identified fluorine-containing and cyclic ether motifs as key structural features for high CE. Accordingly, a fluorine-rich, cyclic-ether-based electrolyte (1 M lithium bis(fluorosulfonyl)imide in methyltetrahydrofuran and 1,1,2,2-tetrafluoroethyl 2,2,3,3-tetrafluoropropyl ether (1:3 v/v)) achieved a CE of 99.72% in Li-Cu cells and excellent cycling stability in lithium-lithium iron phosphate full cells. This framework can facilitate AI-based molecular design for electrochemical energy storage and conversion.

키워드