Integrating physics-based modeling and machine learning for degradation diagnostics of lithium-ion batteries

Traditional lithium-ion (Li-ion) battery state of health (SOH) estimation methodologies that focused on esti-mating present cell capacity do not provide sufficient information to determine the cell's lifecycle stage or value in second-life use. Quantifying the underlying degradation modes that cause capacity fade can give further insight into the electrochemical state of the cell and provide more detailed health information such as the remaining active materials and lithium inventory. However, current physics-based methods for degradation diagnostics require long-term cycling data and are computationally expensive to deploy locally on a device. To improve upon current methods, we propose and extensively test two light-weight physics-informed machine learning methods for online estimating the capacity of a battery cell and diagnosing its primary degradation modes using only limited early-life experimental degradation data. To enable late-life prediction (e.g. > 1.5 years) without the use of late-life experimental data, each of the methods is trained using simulation data from a physics-based half-cell model and early-life (e.g. < 3 months) degradation data obtained from cycling tests. The proposed methods are comprehensively evaluated using data from a long-term (3.5 years) cycling experiment of 16 implantable-grade Li-ion cells cycled under two temperatures and C-rates. Results from a four-fold cross-validation study show that the proposed physics-informed machine learning models are capable of improving the estimation accuracy of cell capacity and the state of three primary degradation modes by over 50% compared to a purely data-driven approach. Additionally, this work provides insights into the role of temperature and C-rate in cell degradation.