Understanding Pitfalls and Opportunities in Estimating Parameters of a Physics-Based Battery Model Using a Machine Learning Based Method—A Case Study With Long Short-Term Memory Neural Network

Abstract

Abstract. Li-ion batteries' design parameters and material properties, such as porosity, electrode thickness, and solid-phase diffusivities, typically vary substantially due to different design goals as well as variations and defects introduced during manufacturing. Many methods have been used for the parametrization of battery cells to enable accurate simulations using physics-based models, including machine learning (ML)-based methods that have become increasingly popular. However, there are inherent limitations to the type and number of parameters that can be estimated using these methods if only standard charge/discharge protocols are utilized. In this work, the typical battery parameters in a continuum-level physics-based model, namely single particle model (SPM), are estimated using the time-series data generated by simulating constant current–constant voltage (CC–CV) charging at different C-rates. These data are first used to train a long short-term memory neural network (LSTM NN) model and then to predict parameters categorized into three groups: electrode design, transport, and kinetics. The parameters are estimated individually (i.e., only one parameter at a time), concurrently (i.e., multiple parameters at a time), and by combining them into one effective parameter. We demonstrate that physics-informed, targeted weighting of selected segments of time-series data, a task for which ML-based parameter estimation is particularly well suited, can substantially improve the identifiability of specific parameters. We find that simultaneous estimation of multiple parameters can yield acceptable agreement in terms of measurable outputs, but the error in internal states must be paid attention to, especially if internal states are to be used for control purposes within a battery management system. We show that discretization errors arising from the numerical solution of partial differential equations can influence model-generated training data. This can ultimately influence the accuracy of the machine learning model which underscores the importance of an appropriately designed grid convergence study.