Investigation of Heat Source Layout Optimization by Using Deep Learning Surrogate Models

Abstract The heat source layout optimization (HSLO) is typically used to facilitate superior heat dissipation in thermal management. However, HSLO is characterized by numerous degrees-of-freedom and complex interrelations between components. Conventional optimization methodologies often exhibit limitations such as high computational demands and diminished efficiency, particularly for complex scenarios. This study demonstrates the application of deep learning surrogate models based on the feedforward neural network (FNN) to optimize heat source layouts. These models provide rapid and precise evaluations, with diminished computational loads and enhanced efficiency of HSLO. The proposed approach integrates coarse and fine search modules to traverse the layout space and pinpoint optimal configurations. Parametric examinations are taken to explore the impact of refinement grades and conductive ratios, which dominate the optimization problem. The pattern changes of the conductive channel have been presented. Moreover, the critical conductive ratio has been found, below which the conductive material can no longer contribute to heat dissipation. The outcomes elucidate the fundamental mechanisms of HSLO, providing valuable insights for thermal management strategies.