Accurate estimation of the state of health (SoH) of lithium-ion batteries is essential for ensuring the safety, longevity, and performance of battery-powered systems, especially in electric vehicles and portable electronics. This study presents a systematic benchmarking of seven open-source automated machine learning (AutoML) frameworks—Fast Lightweight AutoML (FLAML), Tree-Based Pipeline Optimization Tool (TPOT), H2O AutoML, AutoGluon, PyCaret, evaluation machine learning (EvalML), and LazyPredict—applied to the task of battery SoH prediction. Using two public datasets, i.e ., Dataset 1 and Dataset 2, a robust hybrid feature set was extracted for each cell, consisting of equivalent full cycles (EFC), incremental capacity (IC) peak amplitude, and mean temperature. These features capture both short-term operational effects and long-term degradation patterns. The evaluation protocol employed a leave-one-cell-out (LOCO) nested cross-validation strategy, ensuring unbiased generalization assessment across distinct battery cells in two open-source datasets. Among the AutoML frameworks, FLAML, TPOT, and EvalML consistently delivered high accuracy, achieving mean absolute errors (MAE) as low as 0.214% and root mean squared errors (RMSE) as low as 0.264% across multiple cells. Even the worst-performing result (TPOT on Cell 7 with MAE = 0.967%) outperformed state-of-the-art models in the literature, which typically report MAE values between 1.17% and 1.67% using handcrafted IC features. This demonstrates the remarkable capability of AutoML frameworks not only to optimize model architectures and hyperparameters but also to surpass manually tuned traditional approaches. The study confirms that AutoML frameworks such as AutoGluon and H2O AutoML can offer strong generalization with minimal manual intervention, making them highly suitable for real-time deployment in battery management systems (BMS). The results underscore the potential of AutoML in democratizing access to accurate battery diagnostics by balancing predictive performance, computational efficiency, and transparency.
Kadem et al. (Fri,) studied this question.