Data-driven Machine Learning Framework for Life-cycle Prognostics and Degradation Factor Interpretation of Next-generation Lithium Secondary Batteries

Abstract

The accurate prognosis of the remaining useful life of next-generation lithium secondary batteries is a critical requirement for ensuring the reliability and safety of high-power energy storage systems. However, traditional physics-based models often struggle to capture the highly non-linear and stochastic degradation patterns exhibited by lithium-ion and lithium-sulfur batteries, particularly under high C-rate conditions. This dissertation presents a comprehensive data-driven machine learning framework designed for high-precision life-cycle prognosis and the quantitative interpretation of degradation factors. The proposed framework integrates electrochemical differential analysis with advanced ensemble learning techniques to overcome the limitations of conventional diagnostic methods. By utilizing incremental capacity and differential voltage analyses during the early-cycle stabilization window, a robust set of kinetic and thermodynamic descriptors is extracted. These features serve as the primary input for a meta-ensemble voting regressor, which strategically combines the predictive strengths of random forest, gradient boosting, CatBoost, and AdaBoost algorithms. This integration effectively mitigates individual algorithmic biases, resulting in superior R2 performance and a significant reduction in predictive variance across diverse cell capacities ranging from 0.25 Ah to 2.0 Ah. A key contribution of this research is the transition from opaque predictive models to explainable artificial intelligence. By incorporating SHapley Additive exPlanations grounded in cooperative game theory, the framework provides a rigorous mathematical decomposition of model predictions. This allows for the direct mapping of high-dimensional statistical correlations onto fundamental physical degradation modes, such as the loss of lithium inventory and the loss of active material. The analysis successfully identifies the early-cycle precursors of rapid capacity fade in high-rate Li–ion cells and the "sudden death" behavior characteristic of Li–S systems. Ultimately, this work demonstrates that the synthesis of electrochemical domain knowledge and machine learning not only enhances the accuracy of battery life-cycle forecasts but also provides physically meaningful insights into internal decay mechanisms. The developed framework offers a scalable and reliable solution for the real-time monitoring and safety management of next-generation energy storage technologies, bridging the gap between data science and electrochemical engineering.