Combined Multi-Physics and Machine Learning Approach to Forecast SEI Decomposition in LIBs

Over the last two decades, the tremendous efforts of researchers brought the Li-ion battery (LIB), based upon the intercalated graphite electrode, to its advanced level enabling their wide-spread application. Unfortunately, high demand led to many safety problems especially associated with thermal runaway. A very common reason of LIBs’ thermal runaway is the exothermic decomposition of solid electrolyte interface (SEI) leading to a battery’s rapid self-heating and in extreme scenarios to explosion. It is very challenging to suppress, control or sometimes even observe the thermal runaway due to the SEI formation/decomposition in realistic operation conditions. However, the evolution of thermal and electric parameters during LIBs operation may indicate the occurrence of thermal runaway, which will allow its early forecasting and corresponding suppression. In order to predict the thermal runaway in LIBs, we use machine learning (ML) techniques, in particular, convolutional neural network (CNN) and deep neural network (DNN) to identify temperature pattern change leading to the thermal runaway. The ML methods are trained based upon the images obtained from the multi-physics calculations using Comsol software package. The pouch type LIB consisting of five distinct layers, i.e., two current collectors, negative and positive electrodes and electrolyte, is considered. The electrochemical model is based on a 1D continuum description of reaction and transport along electrodes, electrolyte and current collectors, plus an additional dimension in the electrode particle (P2D). 1 The SEI formation/decomposition is modeled using the parasitic current approach. 2,3 We investigate three regimes of thermal runaway occurrence, i.e., a single heat source, two heat sources and multiple sources. Correspondingly, different CNN and DNN architectures have been built and trained based upon the images from multi-physics calculations. The prediction of the trained network has been tested using simulated and literature available data. References [1] Newman, J.; Tiedemann, W. Potential and Current Distribution in Electrochemical Cells. J. Electrochem. Soc. 1993 , 140 (7), 1–5. [2] Wang, K.; Xing, L.; Zhi, H.; Cai, Y.; Yan, Z.; Cai, D.; Zhou, H.; Li, W. High Stability Graphite/Electrolyte Interface Created by a Novel Electrolyte Additive: A Theoretical and Experimental Study. Electrochimica Acta 2018 , 262 , 226–232. https://doi.org/10.1016/j.electacta.2018.01.018 . [3] Safari, M.; Morcrette, M.; Teyssot, A.; Delacourt, C. Multimodal Physics-Based Aging Model for Life Prediction of Li-Ion Batteries. Journal of the Electrochemical Society 2009 , 156 (3), 145–153. https://doi.org/10.1149/1.3043429 .