An investigation of heat transfer of system-on-chip using finite element analysis and optimized graph convolution networks

This study investigates the heat transfer of a system-on-chip (SoC) using finite element (FE) analysis and a graph convolution network (GCN). The computed temperature data are compared and validated using experimental infrared (IR) data. The predicted data successfully capture the heat transfer at interfaces between various components and locations of heat sources within SoC architectures. Parametric studies of the FE simulation are performed to generate a GCN training database. A hyperparameter optimization is implemented to design an accelerated GCN with high predictive performance. In addition, a residual-connected model architecture is incorporated to further increase the accuracy of the model. A comparison among the experimental, computed, and predicted data is performed across the SoC, which indicates that the applied GCN model is very robust and captures the thermal behavior adequately.