Multiphysics-Informed Machine Learning for Battery Design and Health Monitoring

Abstract

Abstract Current Lithium-ion battery (LIBs) designs are nearing the end of their performance capabilities. As the application and demand on these LIBs are growing continuously, there is also a need for continuous innovation both in the area of battery design and the area of battery state of health (SoH) monitoring. Using a silicon (Si) anode instead of a traditionally used graphite electrode allows an increase in the performance of existing LIBs. However, this increased performance comes at the cost of large stresses that develop in the anode as a result of large volumetric changes due to the unique alloying mechanism of the Lithium (Li) into the Si during the battery cycling. These volumetric stresses cause the development of cracks within the Si anode and delamination of the anode from the substrate, which leads to the capacity loss in the battery along with the growth of solid-electrolyte interface (SEI) on the exposed surfaces of the anode. In this study, we develop a physics-informed machine learning (PIML) model to monitor the SoH of the battery based on the output of these coupled failure modes of Si. 3D finite element (FE) models are built to explore how the large volumetric changes cause cracking and delamination in Si and how the SEI growth on the resultant exposed interfaces leads to the degradation of the battery. The outputs of these FE simulations will finally be used to train multiple Gaussian process regression (GPR) models, which combine together to estimate the SoH of the battery under the interactive effects of the studied failure modes. From the results of the study, it is evident that the developed PIML model which uses the output of LIB FE models, which include coupled failure modes, can be used to efficiently (using only 11.4% of charging time) estimate the SoH of the battery with reasonable accuracy (3.04% error).