Deep learning-based reconstruction of supercritical fluids flow field

Supercritical fluids are widely used in heat transfer and energy systems. However, the drastic thermophysical property changes near the pseudo-critical region lead to nonlinear flow and heat transfer behaviors, posing strong challenges for establishing high-efficiency and high-fidelity numerical simulation methods to advance their heat transfer applications. In this study, a multi-module coupled network (MMC-Net) with a tandem structure is proposed based on deep learning for reconstructing the flow field of supercritical hydrocarbon fuels within regenerative cooling channels. To improve the reconstruction accuracy in the entrance region and ensure consistent reliability of performance, a segmented reconstruction method is introduced. The results demonstrate that MMC-Net effectively captures the nonlinear flow and heat transfer characteristics of supercritical hydrocarbon fuels, exhibiting strong extrapolation capability and robustness. Tests on three datasets show that the average relative errors for the temperature and velocity fields are 0.047 and 0.104, respectively. Furthermore, compared to computational fluid dynamics (CFD), MMC-Net reduces computational complexity by approximately five orders of magnitude while still achieving excellent reconstruction of the thermal acceleration phenomenon unique to supercritical fluids. These results prove that the practicality of the network could provide an auxiliary or alternative approach for engineering applications related to supercritical fluids.