• A two-stage topology optimization framework is proposed, which ensures the effective formation of flow channels in complex design domains. • An interactive optimization architecture integrating the SHAP method the ETO algorithm, a CNN, a CNN with channel attention, and a CNN with combined channel and spatial attention was established. • A novel objective function incorporating an area difference correction and a flow field reorganization strategy was proposed. • Compared to traditional straight and bio-inspired channels, it reduces the overall average temperature difference by 43 % and 37 %, and the maximum temperature difference by 48 % and 38 %, respectively, under a lower pressure drop. To address the challenges of low computational efficiency and poor solution quality in topology optimization for turbulent conjugate heat transfer, this study proposes a search-and-shape two-phase optimization framework. The second phase introduces modifications to the objective function to accelerate structural convergence. Furthermore, based on an analysis of conventional initial layouts’ influence on final topologies, we develop an innovative flow field control initialization topology framework named DNNTO. It integrates a reduced-order model combining multi-neural network interactions, the ETO method, and automatic differentiation, effectively balancing search efficiency and computational accuracy in high-dimensional data spaces. A novel multi-objective function based on regional correction and flow field restructuring is also introduced, which uniformly characterizes cooling performance across different solid-phase area fractions. Solving this framework yields a well-structured initial flow field, which is then optimized using an adjoint-based discrete sensitivity model and the GCMMA algorithm. Results demonstrate that the proposed framework achieves superior optimization performance while significantly reducing computational time and structural complexity. Compared to traditional straight and bio-inspired channels, DNNTO reduces the average temperature variation by 43% and 37%, and peak temperature variation by 48% and 38%, respectively, under lower pressure drop. Across various flow conditions, it consistently outperforms conventional designs by at least 80%. When compared to conventional topology optimization on finer meshes, DNNTO reduces optimization time by 80% and structural complexity by 30%, while maintaining superior thermal performance.
Zhou et al. (Fri,) studied this question.