History-aware adaptive mesh refinement via latent-relevance dataset selection for thermal-fluid topology optimization

Topology optimization (TO) of thermal-fluid systems is computationally expensive due to the repeated solution of governing equations. In addition, conventional adaptive mesh refinement (AMR) strategies based on design variables may not accurately capture the variations of physical fields. To address these issues, a latent-relevance dataset selection strategy (LRDS) is proposed to improve the quality of datasets for history-aware field prediction. Based on the resulting high-accuracy predictions, a prediction-guided AMR framework is further developed, in which mesh refinement is driven by the gradients of predicted velocity fields. The proposed framework is evaluated through a multi-inlet flow channel optimization and a convective cooling structure optimization, where the LRDS-enhanced prediction framework achieves average prediction errors of only 0.26% and 1.89%, respectively. Furthermore, the prediction-guided AMR strategy generates meshes that are more consistent with the underlying flow physics and significantly improves mesh quality. The average mesh ratio is reduced to 29.11% and 54.38%, compared with 62.58% and 205.82% obtained using a conventional configuration-based AMR strategy. For the convective cooling case, the average high-fidelity simulation time is reduced by 44.45%, and the total computational time remains 19.60% lower than the uniform-mesh baseline after accounting for the additional LRDS, prediction, and AMR overheads. These results demonstrate that the proposed framework provides an accurate and computationally efficient approach for thermal-fluid TO.