Beyond Predictions: An Interpretable Machine Learning Approach for Battery Performance Forecasting

Abstract

Abstract Lithium-rich layered oxide (LRLO) hold great promise as cathode materials for lithium-ion batteries, but they face challenges due to their complex electrochemical behavior and structural instability. This study proposes an analysis framework using unsupervised learning via Principal Component Analysis (PCA) to improve the predictability and reliability of these materials. By applying PCA, we have identified key factors affecting their electrochemical performance and degradation mechanisms. This has enabled us to easily separate and elucidate oxygen and manganese redox reactions in the low-voltage range, thereby improving our understanding of how the evolution of these reactions affects the degradation of LRLO materials. The PCA-based approach proves to be highly effective in predicting performance and identifying degradation pathways, making a significant advance in the understanding and optimization of these cathodes. These findings represent a step forward in quantifying the mechanisms of electrode materials, which requires the development of models that integrate domain knowledge with data.