Machine Learning-Driven Optimization of Spent Lithium Iron Phosphate Regeneration.

Abstract

The increasing number of spent lithium-ion batteries demands efficient recovery or regeneration to address the associated environmental challenges. Solid-state direct regeneration of spent electrodes is a promising technique that has received significant attention recently. However, the process still requires considerable optimization before being commercially applied. This study leverages machine learning (ML) to develop highly accurate models that characterize the performance of regenerated lithium iron phosphate (LFP) cathodes through three case studies focused on direct regeneration methods. Five different ML models, including artificial neural network (ANN), advanced classification and regression trees (C&RT), boosted regression trees (BRT), support vector machine (SVM), and K-nearest neighbors (KNN), were trained using the collected data. The optimized regeneration conditions identified by the ANN model indicate that a 6.2% increase in specific discharge capacity can be achieved compared to the conditions determined experimentally. The results also showed a possible increase in cycle life, with higher capacity retention after 1147 cycles. These findings highlight the efficacy of ANN models in predicting and optimizing the performance of regenerated batteries, offering significant reductions in time and resources compared to traditional laboratory methods. Moreover, the concept demonstrated in this study shows strong potential for generalization to other battery materials, enabling the optimization of regeneration processes across a broader range of battery chemistry. While most research emphasizes using support vector machines (SVMs) for modeling newly manufactured batteries, this study demonstrates that ANN models provide superior accuracy for regenerated batteries, paving the way for more sustainable energy storage solutions.