EP-BOLFI: Bayesian Optimization for Automated Parameterization of 1+1D Battery Cell Models

Abstract

Our goal is the automated parameterizing of battery cell models for model-based evaluation of experimental databases. The manual standard approach requires cell disassembly and individual measurements on the various cell components. Measurement techniques include, e.g., galvanostatic intermittent titration technique (GITT) or impedance spectroscopy. They are complicated by their long run-time and considerably noise sensitivity. Bayesian algorithms can directly incorporate the inherent uncertainties of model and measurement. The standard approach for parameterization is Markov-Chain Monte Carlo (MCMC). But with 1+1D battery cell models, their simulation time is too large for the tens of thousands of required samples. In this contribution, we extend Bayesian Optimization (BOLFI) with Expectation Propagation (EP) to create a black-box optimizer suitable for modular 1+1D battery cell models. The algorithm can exploit a partitioning of the experimental data into features that is motivated by physico-chemcial understanding. However, the algorithm does not rely on approximative formulas and can be applied to a broad range of techniques. This approach reduces the number of required simulations for four parameters from 100,000 to about 500. Furthermore, we can estimate parameter uncertainties and inter-dependencies. As an example, we process GITT full-cell measurements of lithium-ion batteries to non-destructively characterize the diffusivities of both electrodes at the same time.