Data-driven multifidelity topology design for enhancing turbulent natural convection cooling

This study develops a data-driven multifidelity topology design (MFTD) method for the topology optimization (TO) of extruded fins in a 3D turbulent natural convection system. The high-fidelity simulation resolves the 3D turbulent natural convection using the standard k-ε model combined with the wall function method, while the low-fidelity TO acquires the candidate 2D fin patterns by considering a two-layer Darcy-flow-based natural convection. The framework of evolutional algorithm is employed, and the variational auto-encoder is integrated with the principle component analysis for a more stable and efficient crossover. The proposed framework is applied to design the fins for enhancing the cooling of chips in a 5G active antenna unit considering different settings of heating power. It is found that the hollowed center and the heterogeneous pillar distribution are two significant optimized features in the MFTD designs, which can contribute nearly 30% to the total heat transfer enhancement. In the base case with two 20 W chips, the MFTD design with fin volume fraction as 5% can decrease maximum temperature by more than 10 K and increase the Nusselt number by at least 14.6% compared with the conventional pin fin and plate fin with a similar fin volume fraction, indicating the effectiveness of present design method. Besides, it is noted that the uneven heating power leads to an enhanced structural heterogeneity and slimmer fins, while the lower heating power strongly prohibits the presence of fins with higher fin volume fraction.