Safety problems of high energy density and high-power Li-ion batteries (LIBs) associated with thermal runaway (TRA) have become a serious problem resulting in massive recalls by manufacturers, and at times even endangering the lives of the consumers. In general, the TRA consists of a battery’s rapid self-heating sourced from exothermic (electro-)chemical reactions and/or mechanical abuse. It is impossible to directly monitor the occurring events during the TRA in practical operating conditions (e.g., in cell phones or electric vehicles). However, the change in electrical parameters (pattern) during the TRA-like events could potentially indicate the existence of a failure, thus allowing to foresee the LIBs malfunction.In this contribution, we employ multi-physics modeling and machine learning (ML) to address this important TRA problem in LIBs. Specifically, we use ML techniques to learn and predict the likelihood of the TRA, using the data acquired from the multi-physics modeling of LIBs. The multi-physics modeling approach is comprised of a coupled thermal, electrochemical (P2D model) and degradation sub-models. In this work, a degradation phenomenon leading to the TRA is the solid electrolyte interface (SEI) formation/decomposition in the negative electrode. Subsequently, due to the time-varying characteristic of the variables affecting TRA (i.e, voltage, C-rate, state-of-charge, temperature, etc.), we propose three ML techniques, namely, Support Vector Machine, Deep Neural Network, and Recurrent Neural Network to be customized and applied for determining the TRA likelihood. For deep learning methods, we propose a new activation function, TK-MARS, based upon a partitioning based multivariate adaptive regression spline (MARS). Finally, to mitigate the uncertainty of the prediction results under different degradation conditions, we built an ensemble framework (i.e., a committee of single prediction models) incorporating the trained individual models, in order to maximize the robustness and generalization power of the prediction for a new unseen observation without prior knowledge on the degradation. Thus, the developed combined multi-physics and ML modeling approach establish a basis for accelerated or ‘on-the-fly’ prediction of the TRA as well as a framework for extending machine learning methodologies to broad applications in electrochemistry.