Real-Time Precision Prediction of 3-D Package Thermal Maps via Image-to-Image Translation

Thermal optimization plays a crucial role in the design of advanced packaging technologies. Due to the large number of thermal simulations needed for optimal design, reductions in simulation run-time are critical. Here, we cast the temperature solution process into an image-to-image translation problem. We model the power generation map, conductivity map, and boundary conditions into separate channels of an image. We then generate temperature solutions by training a conditional image generative model, composed of a U-Net shaped generator and a discriminator, using deep neural networks (NNs). The resultant NN model exhibits superior accuracy for unseen inputs. Speed wise, it enables near real-time design, providing a 1663x and 14,885x speedup over a sparse matrix optimized finite element method (FEM) and ABAQUS((R)) respectively.