Predicting battery performance with small datasets by enhancing machine learning with distribution of relaxation times analysis

Accurate battery diagnostic models often struggle in capturing complex degradation mechanisms. Here, we report a combined approach using distribution of relaxation time (DRT) analysis and machine learning (ML) of electrochemical impedance spectroscopy (EIS) data to diagnose and predict the performance of solid-state batteries. The dataset has 112 impedance measurements collected over 100 cycles from eight TDK CeraCharge batteries. Using EIS features and the applied charge-discharge rate, the models estimate battery capacity at a given cycle and predict the capacity after the subsequent ten cycles. ML methods, including Bayesian regression, Gaussian process regression, neural networks, and decision trees, achieve accurate diagnostics (mean absolute error 1.6%) and prediction (2.4%). DRT-based feature selection identifies the most informative frequency range (15 Hz–1.5 kHz) and improves training efficiency. Tomography reveals mechanical degradation, i.e., volume expansion and fractures, which increases impedance and contributes to capacity loss. This approach enables accurate, data-driven battery diagnostics and forecasting.