Early Prediction of Lithium-Ion Battery Cycle Life Using Machine Learning

Accurate early prediction of lithium-ion battery lifetime allows manufacturers for rapid development without the need for long-term aging tests. Instead of cycling cells thousands of times, battery life can be estimated from the first 100 cycles, significantly shortening testing time. Early predictions also give end users practical estimates of remaining battery life, enabling smarter usage, maintenance planning, or replacement decisions. This work presents a machine learning framework that first establishes a LASSO model promoting sparsity, then evaluates four additional methods, support vector machine (SVM), neural network (NN), boosted tree, and Gaussian Process Regression (GPR) models using features extracted from early charge– discharge cycles. Features are based on the variance, discharge, and full feature sets defined in prior work. Unlike previous studies that used fixed training/test splits, we combine all available battery datasets and apply 5-fold cross-validation for robust, unbiased evaluation. Results show that GPR consistently achieves the overall lowest RMSE and MPE across all feature sets. SVM, NN, and boosted tree models also deliver good predictive accuracy when a particular feature set were selected, demonstrating the framework’s effectiveness in capturing degradation trends from limited early-cycle data.