Probing simulation-based inference for credible battery model parameter estimation from impedance data

Electrochemical battery models like the Doyle–Fuller–Newman (DFN) model are increasingly being employed for tasks such as model predictive control during fast charging or model-based cell design. However, their parametrization poses a significant bottleneck for their widespread adoption. The DFN model contains numerous cell-specific parameters, whose determination usually involves time-consuming lab work. Data- driven parameter estimation has therefore emerged as a key factor in battery research. Hitherto, least-squares fitting of model parameters to voltage data via global optimization has arguably been the standard approach in literature. Recently, impedance has shown promise as a more efficient data source, due to its separation of physical processes on different time scales. Additionally, shortcomings of global optimization, e.g., missing information about identifiability, have motivated Bayesian approaches for parameter estimation. Leveraging these advancements, we employ neural posterior estimation, a simulation-based inference algorithm, and compare its performance to particle swarm optimization (PSO), a metaheuristic global optimization approach commonly used in literature, for the estimation of 17 model parameters from synthetic full-cell impedance data with known ground truth parameter values. We use the inferred posterior for an analysis of compensation mechanisms among the parameters of the model and investigate the estimation quality when the prior distribution is sequentially updated. Our findings suggest that Bayesian inference is more efficient and reliable for noninvasive battery model parameter estimation, e.g., yielding an absolute error reduction of around 50% compared to the PSO estimates in the low frequency region, while also enabling insightful analysis of the obtained results. Furthermore, correlations among DFN model parameters are presented and discussed in the context of parametric identifiability for noninvasive parametrization from impedance data. This work hence contributes to the advancement of reliable data-driven parametrization workflows for electrochemical battery models, thereby facilitating model-based research and development across academia and industry.     «

Electrochemical battery models like the Doyle–Fuller–Newman (DFN) model are increasingly being employed for tasks such as model predictive control during fast charging or model-based cell design. However, their parametrization poses a significant bottleneck for their widespread adoption. The DFN model contains numerous cell-specific parameters, whose determination usually involves time-consuming lab work. Data- driven parameter estimation has therefore emerged as a key factor in battery resear...     »