Constructal evolutionary design of liquid cooling heat sink embedded in 3D-IC based on deep neural network prediction

The high degree -of -freedom (DOF) optimized design of the liquid cooling structure embedded in the 3D -IC is critical to fully releasing the cooling potential for heat sinks. For the embedded hybrid heat sink with finned microchannels, this study sets constraints on the total volume of the heat sink and the volume ratio of the internal fins based on the constructal theory. A conditional generative adversarial network (cGAN) was used to predict the physical parameter distribution of the heat sink. A five -degree -of -freedom (5-DOF) evolutionary design of the fin cross-section shape to minimize the maximum temperature of heat sinks was performed by incorporating a genetic algorithm. The trained cGAN had desirable prediction accuracy and much lower computational cost for high-DOF design than the full CFD model. The fin cross section with a larger width and length, and the fins inwardly contract on four sides or produce ribs on four sides, could reduce the maximum and average temperatures of the heat sink. The 5-DOF optimal fin cross-section shape to minimize the maximum temperature of the heat sink is similar to a butterfly wing. During the constructal evolution that minimizes the maximum temperature, the overall thermal performance of the heat sink is also promoted significantly.