Interpretable Learning of Accelerated Aging in Lithium Metal Batteries.

Abstract

Lithium metal batteries (LMBs) with high energy density are perceived as the most promising candidates to enable long-endurance electrified transportation. However, rapid capacity decay and safety hazards have impeded the practical application of LMBs, where the entangled complex degradation pattern remains a major challenge for efficient battery design and engineering. Here, we present an interpretable framework to learn the accelerated aging of LMBs with a comprehensive data space containing 79 cells varying considerably in battery chemistries and cell parameters. Leveraging only data from the first 10 cycles, this framework accurately predicts the knee points where aging starts to accelerate. Leaning on the framework's interpretability, we further elucidate the critical role of the last 10%-depth discharging on LMB aging rate and propose a universal descriptor based solely on early cycle electrochemical data for rapid evaluation of electrolytes. The machine learning insights also motivate the design of a dual-cutoff discharge protocol, which effectively extends the cycle life of LMBs by a factor of up to 2.8.