Physics-Informed CNN for Battery Pack Temperature Fields

Researchers at the University of Illinois Urbana-Champaign built a 21700 battery pack with indirect liquid cooling (cold plates above and below, thermal paste around the cells), calibrated a simplified finite-element model against experiment, then trained a physics-informed convolutional neural network to predict the temperature field.

How physics enters: The loss function is constructed from the heat conduction equation discretised by finite differences, so the network is penalised for violating energy conservation rather than only for pixel-wise error. This also stabilises training when labelled data are scarce.

Reported results: The physics-informed network improves accuracy by more than 15% over the data-driven baseline trained on exactly the same dataset.

Industrial relevance: Battery thermal management design is dominated by the cost of generating finite-element training data. Embedding the governing PDE lets a surrogate reach usable accuracy with far fewer simulations - the same pattern that makes physics-informed methods attractive across CAE.