High-fidelity three-dimensional aerodynamic flow prediction on wings with physics-constrained dual-parallel attention UNet++

Accurate flow field data provide a robust foundation for analyzing concentrated force distribution and implementing flow control strategies. Nevertheless, current deep neural network methods exhibit limitations in accuracy when applied to reconstruct three-dimensional wing flow fields. To address this challenge, we propose an intelligent flow field reconstruction technique termed physics-constrained Dual-Parallel Attention UNet++ (DPAtt-UNet++). This method utilizes the Unet++ neural network architecture as its backbone, integrating a dual-parallel attention mechanism and nested network structure. Furthermore, a physics-constrained hierarchical loss function is introduced, incorporating the residuals of the governing Navier-Stokes equations as soft constraints to enforce physical consistency during training. Comprehensive evaluations demonstrate that the proposed DPAtt-UNet++ outperforms not only the baseline U-Net by approximately 10% in reconstruction accuracy, but also shows clear improvements over both standard UNet++ and a non-physics-constrained DPAtt-UNet++, validating the effectiveness of the integrated attention mechanism and physical constraints. Tests on wings constructed from different airfoil profiles confirm robust generalization capability across varying flow conditions and geometric shapes. Moreover, the method achieves approximately 2–3 orders of magnitude faster reconstruction speed compared to the Computational Fluid Dynamics (CFD) method in the online prediction phase. These results demonstrate the method can accurately and efficiently reconstruct flow fields for different geometries under various flow conditions.