A physics-driven generative model to accelerate artificial intelligence development for lithium-ion battery diagnostics

Recent advancements in artificial intelligence (AI) have highlighted the potential of extensive battery cycle data for diagnosing performance degradation in lithium-ion batteries (LIBs). Therefore, related generative models have been developed to rapidly generate a variety of battery data from a small number of real-world experimental measurements. This study proposes a physics-driven generative model (PGM) to produce realistic battery cycle data by reasonably sampling electrochemical parameters in a stochastic manner. PGM captured the fundamental principles of LIBs as electrochemical parameter distributions and then generated the corresponding virtual LIBs, even under conditions that have not been previously applied. This superior generalizable capability was validated by showing that PGM was very effective in detecting the internal short circuits (ISCs) of LIBs. Synthetic data with different ISC degree effects were generated using PGM based only on ISC-free LIBs. The results showed that a neural network trained on the generated synthetic data achieved a detection accuracy of 97.39 % for real physical ISCs, which was comparable to the detection results of a neural network trained using approximately 25 times more real data, including experimental results with ISCs. The proposed PGM is expected to contribute significantly to the rapid advancement of AI-based LIB diagnostics by generating physically meaningful data based on internal electrochemical principles.