Clustering and Prediction of Early Capacity Fade Trajectories in Lithium-Metal Batteries using Data-Driven Approaches

Lithium-metal batteries (LMBs) are promising candidates for next-generation batteries due to their high energy density. However, during cycling, lithium forms dendritic structures on the negative electrode surface, leading to the formation of dead lithium and resulting in capacity loss. In this paper, we employ a dataset of lab-scale lithium-metal batteries cycled under various charging rates and fixed discharge rate. We apply machine learning models that leverage initial cycle data from these cells to predict the state-of-health (SOH) trajectories. The data reveals different forms of SOH trajectories, that indicate different modes of degradation. Through feature engineering, we identify which features most effectively cluster the SOH trajectories of the cells into distinct groups from as low as 10 cycles of initial cycling data. Each cluster is then used to train both a long-short term memory (LSTM) and a gaussian process regression (GPR) model to predict the remaining degradation trajectory (for example, after cycle 10). Overall, we find that the models predict the trajectory within a 3% mean absolute percentage error (MAPE) margin, with the GPR model delivering twice the accuracy of the LSTM model in predicting capacity trajectories.