A Combined Multi-Physics Modelling and Machine Learning to Predict Electro-Thermal Failures of Cylindrical Li-Ion Batteries

Safety aspects of Li-ion batteries (LIBs) operation become increasingly important due to their integration into large-scale systems such as electric vehicles (EVs). Recent events involving many EVs explosions endangering consumers’ lives have shown that electro-thermal aspects of LIBs are far from being adequately managed. Thus, one of the most crucial goals of current research is to steer LIBs’ technology toward long-lasting and safe operation. In this work, we employ data-driven methods to predict and potentially prevent the thermal runaway (TRA) in LIBs. Specifically, we use various machine learning (ML) techniques to learn and predict the likelihood of the TRA, using thermal images acquired from the multi-physics modeling of LIBs. The multi-physics modeling approach is comprised of a coupled thermal, electrochemical (P2D) model1, and degradation due to the solid electrolyte interface (SEI) formation/decomposition sub-models.2 The presented multi-physics model has been formulated as a three-dimensional model to facilitate efficient and accurate numerical simulations in different conditions. Under general working conditions, the cell is exposed to a continuous charge/discharge load, where heat is generated inside the cell due to several electrochemical and degradation processes. Thermal images of the battery surface are collected as sequences of video frames during the battery operation from multiple simulations to be used in ML. LIB’s geometrical and material properties are taken from the 2170 cell as used in the current Tesla vehicles (model S 2021, model 3 2021). Based upon the multi-physics modeling results, two deep learning techniques, i.e., Convolutional Neural Networks (CNN) and object detection using YOLO, are proposed to address the problem of predicting the TRA like-events.3–5 The multi-physics modeling results in the form of the temperature profiles of the battery surface as well as the SEI formation/decomposition rate at various charge/discharge cycles are validated against the prior literature experimental data. Two regimes with single and double hot spots in the cylindrical LIB are investigated for a potential TRA occurrence. The CNN predictions of the TRA-like event on unknown data set show an excellent accuracy reaching up to 90% for all the different CNN architectures used. The object detection technique is found to be very efficient in identifying the location of TRA initiation at various conditions. Thus, the developed combined multi-physics and ML modeling approach establishes a basis for 'on-the-fly' prediction of the TRA as well as a framework for extending machine learning methodologies to broad applications in electrochemistry. References J. Newman and W. Tiedemann, J. Electrochem. Soc., 140, 1–5 (1993). X. Fen, X. He, M. Ouyang, L. Wang, L. Lu, D. Ren and S. Santhanagopalan, J. Electrochem. Soc., 165, A3748–A3765 (2018). J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., 2016-Decem, 779–788 (2016). K. He, X. Zhang, S. Ren, and J. Sun, Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., 2016-Decem, 770–778 (2016). K. Simonyan and A. Zisserman, 3rd Int. Conf. Learn. Represent. ICLR 2015 - Conf. Track Proc., 1–14 (2015).