Physics-informed neural networks for thermal spreading of multilayer multichip power modules

Efficient thermal management in multilayer, multi-source power modules remains a critical barrier to advancing high-performance electronic systems. Analytical models are limited to highly simplified geometries, while finite volume method (FVM) requires extensive meshing and prohibitive computational cost for iterative design exploration. Here, we present a physics-informed neural network (PINN) integrated with a Gaussian process (GP) surrogate to enable mesh-free, physically consistent thermal prediction and rapid design optimization. The PINN achieves prediction errors below 2.4% for single-source validation and below 2.0% for a realistic seven-layer, multi-source power module, while the complete PINN-GP workflow reduced the estimated design-exploration time from approximately 15,000 to 200 h compared with sequential FVM evaluation. Coupling the PINN with a GP surrogate enables efficient identification of optimal geometric parameters, achieving up to 4.1% reduction in thermal spreading resistance. This PINN-GP framework provides a scalable pathway for thermal analysis and design optimization in next-generation power modules.