Spatial structure-embedded physics learning for spatiotemporal modeling of large-format pouch lithium-ion battery

Large-format pouch lithium-ion batteries exhibit complex and highly nonlinear thermal behaviors both spatially and temporally. Accurately modeling these behaviors is critical for ensuring their optimal performance, longevity, and safety. Traditional physics-based models have strong generalization but struggle to address system uncertainties in real-world applications. Data-driven methods can capture these uncertainties but rely on extensive sensor data. To address these challenges, we propose a spatial structure-embedded physics learning framework to model the complex thermal behaviors of large-format pouch lithium-ion batteries. In this framework, the electrochemical-thermal mechanisms are used to guide the network training process. A set of neural spatial basis functions that represent spatial evolution characteristics of thermal equations are designed and embedded into the network. This integration enhances the model’s physical fidelity, enabling accurate temperature prediction across large computation domains. Additionally, an improved automatic differentiation scheme that utilizes the predefined SBFs is developed to calculate derivatives in the thermal equations, thereby reducing network’s training cost. Using only a few boundary data to calibrate the physical equations, our model exhibits superior performance across various operating conditions. Results demonstrate that the proposed method achieves lower RMSE values, with a maximum of 0.154 under 5 C condition and 0.026 under UDDS condition. Additionally, the method requires 36.1% fewer trainable hyperparameters compared to the original physics-informed machine learning method, leading to a reduction in total training time by approximately 45.6%. © 2026 Elsevier Ltd.