Triply Periodic Minimal Surface (TPMS) structures have emerged as a promising solution for next-generation high-efficiency heat exchangers due to their high specific surface area, lightweight characteristics, and smoothly connected internal channels. However, efforts to enhance heat transfer performance often lead to increased flow resistance, necessitating structural optimization. The complex geometry of TPMS structures also results in high computational costs for numerical simulations, making machine learning-based surrogate models a practical approach for efficient design. To address the challenges of high simulation costs and limited training data, a voxel-based geometric segmentation and data augmentation method was developed. A deep neural network surrogate model was constructed and trained, achieving an R 2 value exceeding 0.96. This model was applied to guide the geometric design of novel hierarchical TPMS structures, reducing the simulation time from 74 minutes to 9 minutes. Numerical results indicate that compared to the initial single-layer design, the hierarchical Gyroid structures improve heat transfer efficiency by 26.32%–66.39% and yield a more uniform temperature distribution. Further optimization using the trained model reveals that the heat transfer coefficient can be increased by 83.49%, while the pressure drop can be reduced by 72.97%. These findings offer new insights into the design and optimization of TPMS-based heat exchangers for advanced thermal management applications.
Xue et al. (Sun,) studied this question.