Physics-based inverse modeling of battery degradation with Bayesian methods.

Abstract

To further improve Lithium-ion batteries (LiBs), a profound understanding of complex battery processes is crucial. Physical models offer understanding but are difficult to validate and parameterize. Therefore, automated machine-learning methods (ML) are necessary to evaluate models with experimental data. Bayesian methods, e.g., Expectation Propagation + Bayesian Optimization for Likelihood-Free Inference (EP-BOLFI), stand out as they capture uncertainties in models and data while granting meaningful parameterization. An important topic is prolonging battery lifetime, which is limited by degradation, such as the solid-electrolyte interphase (SEI) growth. As a case study, we apply EP-BOLFI to parametrize SEI growth models with synthetic and real degradation data. EP-BOLFI allows for incorporating human expertise in the form of suitable feature selection, which improves the parametrization. We show that even under impeded conditions, we achieve correct parameterization with reasonable uncertainty quantification, needing less computational effort than standard Markov chain Monte Carlo methods. Additionally, the physically reliable summary statistics show if parameters are strongly correlated and not unambiguously identifiable. Further, we investigate Bayesian Alternately Subsampled Quadrature (BASQ), which calculates model probabilities, to confirm electron diffusion as the best theoretical model to describe SEI growth during battery storage.