A Flexible Approach To Parameter Estimation Of Physics-Based Battery Models With Bayesian Optimization

Abstract

A wealth of measurement techniques are available for determining transport or thermodynamic properties of batteries. Some examples are GITT, pulse experiments or least squares with impedance spectroscopy, which are excellent at automatically retrieving a subset of model parameters with relatively low computational cost. They achieve this at the cost of accuracy and compatibility, since each employs a different approximation in order to obtain analytic expressions. So it remains a challenge to study the interactions and effects of multiple model parameters at once, since various measurement techniques would need to be combined intoone so-called cost functional. This cost functional is a function of the measurement data and the battery model parameters and can be used to determine the parameters that best describe the data. The automated computation for this usually leverages the derivative/gradient of the cost functional by the parameters. Beyond some degree of complexity, this gradient would become intractable or unfeasible to analytically calculate, making the computation expensive and/or numerically unstable.We want to eliminate the need for those gradients. The goal is to enable automated material screening with flexible selection of various measurements. Our approach uses the existing specialized techniques as a starting point and refines their results by performing simulations with randomized parameters and comparing them with allavailable measurements. This Monte Carlo approach is made computationally feasible by utilizing Bayesian Optimization for sparse sampling and preprocessing the measurements into intuitive features. The latter also enables us to use measurements that are most often fitted by hand, e.g. discharge curves. Additionally, we make use of a single particlemodel with electrolyte from and/or PyBaMM to make the simulations as efficientas possible, but any 1D+1D model can be used thanks to the sparse sampling approach.