Dual-attention FlowNet: Combining convolutional block attention module and transformer for flow fields reconstruction in supercritical fluids

Due to significant thermophysical property variations and intricate vortex structures, traditional numerical simulations of supercritical fluids (SCFs) are computationally intensive. This study proposes Dual-Attention FlowNet (DAF-Net), a hybrid framework combining a Transformer encoder with a Convolutional Block Attention Module (CBAM) to resolve these complexities. The architecture leverages Transformer-based global dependencies to replicate nonlinear thermal acceleration, while the CBAM enhances localized feature extraction for precise identification of structures such as Dean vortices. To ensure rigorous evaluation, a case-based data splitting strategy is employed, validating the model's generalization across the operational parameter space. Comparative benchmarking against state-of-the-art (SOTA) models, including Senseiver and MMC-Net, demonstrates that DAF-Net exhibits superior structural fidelity and noise suppression, particularly in high-gradient regions where traditional architectures often suffer from numerical artifacts. Notably, DAF-Net maintains high reconstruction accuracy without requiring the segmented training protocols typically necessitated by sharp property gradients. With a structural similarity index (SSIM) of 0.991 and a peak signal-to-noise ratio (PSNR) of 54.86 dB, this framework provides a robust and high-fidelity surrogate for real-time SCFs analysis. The proposed approach establishes a new paradigm for digital-twin applications in aerospace and energy thermal management systems.