With the rapid development and widespread applications of lithium-ion batteries (LIBs), there is an ongoing need to extend and apply theoretical models that assist LIB’s safety aspects. It is particularly important for electric vehicles (EVs) due to numerous recent fire accidents. Thermal runaway (TRA) is one of the principal causes of LIB’s failures in EVs occurring due to thermal or mechanical breakdown, internal/external short-circuiting, or electrochemical abuse. During EV’s operation, it is impossible to directly monitor the TRA; however, the change in thermo-electrical characteristics (pattern) during TRA-like events could signal the presence of a failure, allowing for the prediction of LIB malfunction. Thus, in this work, we employ machine learning-based techniques informed by multi-physics models to predict and prevent the TRA in large pouch LIBs as presently used in various EVs.

The multi-physics model is implemented in commercial software Comsol, with the P2D electrochemical model¹ and a 3D thermal model. The degradation sub-model² includes oxygen release in the positive electrode to simulate the overcharge phenomenon during EV’s charging. In addition, the oxygen released in the positive electrode may exothermically react with the electrolyte as well as create significant stress in the electrode, which may lead to the mechanical deformation of the electrode and a subsequent TRA. An LG Chem lithium-ion pouch cell consisting of Li(Ni_0.6Mn_0.2Co_0.2)O₂ – NMC622 – cathode and graphite anode are studied to address this severe TRA problem.

As a result of the time-varying nature of the variables that affect TRA, we propose three potential machine learning algorithms. These are Support Vector Machine, Deep Neural Network, and Recurrent Neural Network, tailored and implemented for estimating the TRA likelihood, using thermal images acquired from the multi-physics modeling of LIB pouch cells. Hyperparameters optimization has been performed to identify a set of variables for the best performing ML method. The proposed combined multi-physics and machine learning modeling methodology provide interesting insight and predictive capabilities for TRA prediction.

References

1. J. Newman and W. Tiedemann, J. Electrochem. Soc., 140, 1–5 (1993).
2. X. Feng, X. He, M. Ouyang, L. Wang, L. Lu, D. Ren and S. Santhanagopalan, J. Electrochem. Soc., 165, A3748–A3765 (2018).