A Comparative Study on Physics-Informed Machine Learning for Battery Degradation Diagnostics

Abstract

Quantifying the extent of degradation in lithium-ion battery cells using non-destructive approaches can provide valuable insights into the cell’s state of health and remaining useful life. However, physics-based degradation diagnostic methods typically require collecting long-term aging data and are computationally expensive to deploy locally on a device. This work investigates combining physics-based modeling and machine learning to retain high diagnostic accuracy while mitigating the need to collect long-term degradation data. To transfer knowledge from a physics-based model to a machine learning model, we develop two new approaches: co-kriging and physics-informed neural networks. We aim to effectively diagnose cell health in the late aging stage without needing long-term degradation data. Using a small set of early-life experimental data collected from an aging experiment, we train a co-kriging machine learning model to learn the difference between the experimental data and a set of simulation data generated from a physics-based half-cell model. Additionally, we train a neural network model using a physics-informed loss function with a new term that penalizes discrepancies between neural network model prediction and physics-based model prediction. The trained models can then be used to diagnose cell capacity and three primary component-level degradation modes at a future time. The proposed methods and two existing physics-informed machine learning methods named data augmentation and delta learning are comprehensively evaluated using data from a long-term (4+ years) cycle aging experiment of 24 implantable-grade Li-ion cells cycled under two different temperatures and C-rates. Cross-validated results show that the proposed physics-informed machine learning models can significantly improve the estimation accuracy of cell capacity and three primary degradation modes compared to a purely data-driven approach. Furthermore, this work comprehensively analyzes the trends of the degradation modes and provides insight into the long-term performance of high-quality implantable-grade lithium-ion cells.