Physics-informed latent ensemble DeepONet with synthetic data for real-time thermal runaway prognosis of a lithium-ion battery

Abstract

Real-time prognosis of thermal runaway (TR) is critical for enabling proactive and safe battery thermal management systems (BTMSs), yet it remains challenging due to the complex multiphysics nature of TR. This study proposes a novel physics-informed latent ensemble deep operator network (PILE-DeepONet) trained with synthetic data, aiming to predict TR of a lithium-ion battery using only prior surface temperature in real-time. The framework introduces four key contributions toward a control-enabling solution for BTMSs. First, synthetic data are generated through multiphysics finite element analysis. This feature provides PILE-DeepONet with a virtual sensing capability. Second, the proposed neural network is co-trained with physical intelligence. Physical intelligence enforces physical consistency, improving reliability and generalization capabilities under diverse thermal abuse conditions. Third, two surrogate neural networks are designed to effectively extract and ensemble latent features from prior surface temperature, which inherently contains limited information in real-time measurements. Fourth, novel training strategies are employed to reduce the training complexity associated with physical intelligence co-training. These strategies not only enhance the expressiveness of temporal and spatial gradients but also balance multiple losses, improving convergence stability during training. Quantitative comparisons and ablation studies validate the contribution of each key feature, confirming the superiority of the novel surrogate architecture and training strategies in addressing failure challenges in conventional neural networks and physics-informed neural networks. A feasibility study using noisy synthetic data for uncertainty quantification provides practical guidance for defining safety margins. Feasibility is also supported by validation with an actual experiment, which also demonstrates the virtual sensing capability of future cell states 900 s in advance with an inference time of 4.96 ms. The generality of the proposed neural network is further demonstrated by accurate TR prediction in both LFP cylindrical and NMC pouch cells. PILE-DeepONet enables real-time TR prognosis, offering a promising pathway toward artificial intelligence transformation in BTMSs to ensure the safety and efficiency of lithium-ion batteries.