Key points are not available for this paper at this time.
Battery health often evolves from an initial regime of slow capacity decay to a fast decay regime, due to catastrophic electrochemical reactions to cause the battery failure. The signals of such failure, however, often emerge much earlier during cycling, an early detection of which may help prevent the failure through real-time adjustment of cycling procedures. Here we demonstrate the integration of machine learning module directly into electrochemical battery test stations to constantly detect the signals from battery current and voltage profiles, followed by reinforcement learning–based real-time adjustment of cycling procedures for greatly elongated battery lifetime. We use solid-state lithium metal batteries to demonstrate this proof-of-concept implementation of built-in intelligence for future battery management systems, leveraging their high current density and fast cycling to generate abundant data and greatly shorten the development feedback loop. This intelligent approach improves battery lifetime by 265% and accumulative specific energy based on active material by 250% at 80% state of health, which is achieved through an effective control of detrimental interface reactions. This work showcases a pathway toward real-time artificial intelligence control of batteries for longevity and safety beyond human expertise. Battery health can rapidly decline after subtle early warning signs. Here the authors embed machine learning into test stations to read voltage–current profiles for state perception and adjusting charging strategies, extending solid-state battery life by suppressing harmful interface reactions.
Wang et al. (Tue,) studied this question.