Hybrid loss-based convolutional encoder–decoder neural networks for the robust topology optimization of porous heat exchangers

Deterministic topology optimization (DTO) for cooling microchannels has been widely studied in recent years. However, for the robust topology optimization (RTO) of comprehensive heat transfer performance under the condition of extensive random uncertainties in boundary conditions and material properties, no research has been conducted due to the high computational cost of conjugate heat transfer simulations in complex topological structures. To address this research gap, this study establishes a fluid-thermal RTO model for porous heat exchangers, where the random uncertainties of porosity and inlet velocity are considered. For solving the RTO model, an adaptive convolutional encoder–decoder neural network (CNN) approach is developed by proposing a hybrid data-physics-informed loss function and adaptive updating strategy to capture the topological characteristic features in each RTO iteration and reconstruct the primal physical fields corresponding to the specific realization of random porosity and inlet velocity. The results of solving both DTO and RTO demonstrate that the proposed adaptive updating strategy of CNN can maintain both the hybrid loss function and absolute prediction errors within acceptable thresholds. When the flow dissipation constraints are, respectively, set to 2, 10, and 20 times that of conventional straight single-channel designs, the RTO configurations can achieve reduction in average temperature mean of 11.25%, 14.48%, and 16.88%, along with the decrease in average temperature standard deviation of 31.02%, 27.08%, and 25.15% compared to DTO configurations, respectively. Furthermore, as the inlet velocity mean is increased from 0.05 to 0.30 m/s, the RTO configuration exhibits excellent potential in both mean and robustness of heat exchange performance, and it achieves the relative reduction of 12.26%–15.62% and 12.35%–31.04% in mean and standard deviation of average temperature compared to the DTO configuration, respectively.