Physics-Guided Descriptors Enable Data-Efficient Prediction of Battery Coulombic Efficiency

Lithium metal batteries promise energy densities beyond 500 Wh kg^-1; but their practical deployment remains limited by low Coulombic efficiency and uncontrolled electrolyte-interface reactions. Here, we show that physics-guided machine learning can identify the molecular origin of Coulombic efficiency (CE) from small experimental datasets by embedding 3D electrolyte structures into data-driven descriptors. Among the descriptors examined, the physics-derived solvent-surrounding-Li⁺ descriptor (Li^SSL) enables accurate CE prediction, achieving a test-set R² of 91.15%. Explainable machine learning further reveals Li^SSL as the dominant factor governing model performance, indicating that high-efficiency lithium deposition requires suppression of direct Li^+-solvent interactions. This insight establishes a molecular design principle for electrolytes: weakening solvent participation in the primary Li⁺ solvation environment promotes higher CE. Our work provides a physics-informed, data-driven framework for accelerating electrolyte discovery toward high-energy lithium metal batteries.