Machine Learning Strategies for Forecasting Discharge Capacity in Lithium-Ion Batteries with NCM Layered Cathode Material

The nuanced grasp of determinants impacting discharge capacity stands as an imperative linchpin for propelling the evolution of lithium-ion batteries to new frontiers. We have curated a dataset comprising data from 147 sets of lithium-ion battery cathodes, scrupulously extracted from pertinent literature. Initially, the relationships between variables were visually depicted through Pearson correlation coefficient plots. Subsequently, six models were employed for data prediction. Notably, the gradient boosting model exhibited superior performance, yielding minimal root mean square errors for initial discharge capacity and post-cycling discharge capacity at 12.58 mAh∙g-1 and 15.00 mAh∙g-1, respectively. Subsequently, we conducted an analysis of feature importance and Shapley additive explanations (SHAP) plots to identify the primary factors influencing both initial discharge capacity and discharge capacity after cycling. Among them, for the IC, the highest importance scores are assigned to current density and maximum cyclic voltage, standing at 0.225 and 0.154, respectively. This implies that, compared to other descriptors, current density and maximum cyclic voltage exert a more substantial influence on the GBM model. In the case of the EC, the importance score for the IC significantly surpasses others, reaching 0.353, showcasing a heightened level of contribution to the model. This analysis offers valuable insights guiding the subsequent exploration of conditions impacting discharge efficiency in Li[NixCoyMn1–x–y]O2 (NCM) cathode materials.