Physics-Informed Neural Networks for Degradation Diagnostics of Lithium-Ion Batteries

Abstract Quantifying the extent of degradation in a lithium-ion battery cell can provide valuable insights into the cell’s state of health. However, physics-based methods capable of diagnosing component-level degradation typically require long-term aging data and are computationally expensive to deploy. This work investigates combining physics-based modeling and data-driven machine learning to retain high diagnostic accuracy while mitigating the need for long-term degradation data. We develop a physics-informed neural network (PINN) algorithm to diagnose cell health in the late aging stage without needing long-term data. Using a small set of early-life experimental data from an aging experiment, we train a neural network using a physics-informed loss function which penalizes discrepancies between the neural network’s and the physics-based model’s estimates of four cell health parameters. The trained network can then estimate cell capacity and diagnose three primary degradation modes in the late-aging stage. The proposed method is evaluated and compared with other machine learning algorithms using data from a long-term (3.5 years) cycling experiment of 16 implantable-grade lithium-ion cells. Cross-validation results show that the proposed PINN algorithm can improve the estimation accuracy of the health parameters compared to a purely data-driven approach and has comparable accuracy to the other two physics-informed machine learning methods.