Data-Driven Lithium Salt Design for Long-Cycle Lithium Metal Battery

Abstract

The development of high-performance electrolytes for lithium metal batteries (LMBs) is hindered by time-intensive experimental characterization. Here, a data-driven predictive model is presented to estimate Coulombic efficiency (CE) and lithium metal thickness evolution using electrolyte composition and density functional theory (DFT)-derived descriptors. By analyzing 21 lithium salts, key computational parameters, including LUMO energy levels, are extracted from lithium oxidation states, and adsorption energies. Machine learning models, particularly XGBoost and random forest, achieve high predictive accuracy, reducing mean squared error by over 50% compared to structural-only models. Linear regression reveals that higher LUMO values and lower lithium oxidation states correlate with improved CE, guiding the selection of LiDFP, LiNO₃, LiPDI, and LiHDI as promising additives to LiFSI. While constrained by limited SEI characterization and dataset size, this study establishes a computational framework for electrolyte optimization, accelerating LMB development and cycle life enhancement.