Deep learning to predict battery voltage behavior after uncertain cycling-induced degradation

Predicting battery degradation is not only beneficial for the early detection of risks but also essential for advancing research on energy chemistries. Battery degradation is highly dependent on cyclic conditions, and leads to degradation in its capacity, internal resistance, polarization, etc. Despite decades-long efforts, existing predictions are only valid for one or a few degradation components after a fixed cycling-induced degradation, providing very limited insights. Here, we extend the prediction to the exhaustive voltage responses to various current profiles after uncertain cycling-induced degradation. This is achieved by developing a novel deep-learning framework that integrates future cyclic conditions and a battery model. A long short-term memory network is designed to summarize any given variable-length future cyclic condition flexibly for prediction. The battery model ensures that predictions generalize to various current profiles. As a case study, a large dataset covering 29,760 degradation cycles from 43 lithium-ion batteries is generated under uncertain cyclic conditions for validation. The results show that, by only using data from the initial cycle, the proposed framework can achieve accurate prediction with a median root mean square error below 12.7 mV over the entire state of charge range. This work opens up a promising avenue for conventional battery models to bridge future voltage behaviors of batteries, which is extremely helpful in understanding the degradation of electrochemical energy storage devices.