Adaptive mask physics-driven neural network for complex flow field prediction using self-supervised learning

Abstract

High-precision complex flow field prediction faces challenges in balancing accuracy, efficiency, and physical consistency. This study proposes an Adaptive coding Multi-airfoil Physics-Driven Neural Network (AMPD-NN) employing variable wall condition processing technology for unified prediction across airfoil families with thickness ratios of 6%-15% and camber variations of 0%-4%. The model uses self-supervised training based on physics conservation equation residual minimization, requiring only boundary conditions and geometric parameters while reducing labeled data dependence. Under subsonic conditions, flow field prediction errors remain within 4% with SSIM of 0.965–0.973. Lift and drag coefficient prediction errors are 2.22%-2.70% and below 5%, respectively. Transonic extrapolation validation shows increased errors of 15–16% but within acceptable ranges. Compared with ANSYS Fluent, the model achieves 78–109 × speedup for single predictions and 2.5–2.7 × improvement for batch processing while eliminating mesh generation, enabling efficient parametric analysis.