Physics-Informed Neural Networks for Predicting Thermal Runaway Trajectories across the Full Life Cycle

Thermal runaway is one of the most critical safety challenges in the deployment of lithium-ion batteries across electric vehicles and stationary energy-storage systems. Battery aging process presents a significant influence in thermal runaway behavior due to side reactions such as lithium plating and solid-electrolyte interphase (SEI) formation. Therefore, accurately predicting the onset and evolution of thermal runaway throughout a battery’s entire life cycle is essential for designing safer systems. However, existing work mainly focuses on predicting the behavior of fresh cells and generally requires a large amount of experimental data. In addition, most existing prediction models are strongly tied to the specific testing boundary conditions. For example, results obtained from an Accelerating Rate Calorimeter (ARC) and a Copper Plate Battery Calorimeter (CPBC) often exhibit markedly different thermal processes, which may lead to inconsistent or biased predictions when extrapolated beyond their respective testing environments. To address these gaps, this study proposes a physics-informed neural network (PINN) framework capable of predicting thermal runaway trajectories of NCM/Gr batteries across the full life cycle of lithium-ion batteries. In PINN framework, by embedding energy conservation laws and thermochemical equations into the loss function, the model simultaneously learns from sparse experimental observations and adheres to governing physical constraints. This enables the robustness of prediction on various thermal boundary conditions ( i.e., ARC and CPBC), such that benefits its application across different thermal abuse scenarios. The chemical reaction neural network (CRNN) is incorporated as a sub-solver to model reaction kinetics and heat generation associated with thermal runaway. The CRNN enables the representation of multiple side reactions, including those induced by aging. In this study, the lithium–electrolyte reaction is introduced as an additional aging-related side reaction. This reaction generates extra heat at an early stage, thereby accelerating the temperature rise and leading to an earlier onset of thermal runaway. Our results show that the PINN framework accurately reconstructs the full thermal runaway trajectory, from the initial heating stage to the onset of runaway and the subsequent cooling process. The model successfully predicts thermal runaway behavior across different aging conditions. Furthermore, the CRNN component quantifies the influence of degradation mechanisms, such as SEI decomposition, lithium–electrolyte reactions, and anode–electrolyte decomposition, on the thermal runaway process. The corresponding heat-generation profiles are also extracted, enabling a more comprehensive evaluation of the associated thermal hazards. Overall, this work establishes a PINN-based modeling framework integrated with a CRNN architecture for predicting thermal runaway across the entire battery life cycle. The results demonstrate the strong potential of physics-informed machine learning to overcome experimental limitations, improve predictive robustness, and support the safety evaluation of next-generation lithium-ion battery technologies.