Machine-Learning Prediction of Lithium-Ion Battery Cycling Stability from Cathode Structural and Interface-Relevant Descriptors

Achieving high usable capacity in lithium-ion batteries is valuable only when cathode-related structural and electrode/electrolyte instability can be limited during cycling. This study develops a two-stage TabNet-XGBoost framework to predict capacity retention and cycle life from a cathode-centered, cell-contextual descriptor set spanning cathode composition and crystal structure, anode and electrolyte variables, initial electrochemical characteristics, and cycling conditions. TabNet generates 64-dimensional latent representations for XGBoost regression, while a separate descriptor-level XGBoost model supports SHAP analysis without assigning latent-feature attributions to individual physical descriptors. On a held-out test partition containing only real records, the primary model achieved R² values of 0.980 and 0.982 and RMSE values of 0.022 and 46.2 cycles for capacity retention and cycle life, respectively. A SHAP-informed particle-swarm search identified a candidate region with model-predicted capacity retention of 89.2% and cycle life of 1,419 cycles, compared with reference predictions of 84.6% and 1,185 cycles. These predictions and attributions describe associations within the represented chemistries and conditions; they do not resolve specific bulk or interfacial degradation mechanisms. The candidate is likewise a surrogate-model result that requires synthesis and controlled cycling validation. The framework therefore provides a data-driven screening tool for relating cathode structure and interface-relevant variables to cycling-stability outcomes while defining explicit boundaries for mechanistic interpretation and transfer to unseen systems.