A fast prediction approach for airfoil flow field based on the deep neural network framework

Abstract

A rapid aerodynamic prediction framework, termed ParaAero-Net, is proposed for low-Mach-number two-dimensional airfoils under small-sample conditions. Unlike conventional data-driven approaches that focus only on either flow-field reconstruction or aerodynamic coefficient prediction, ParaAero-Net integrates geometry-aware full-field reconstruction and wall aerodynamic quantity prediction within a unified framework. The first module, Physics-informed U-Net Attention module (PIUA), is developed for flow-field prediction by integrating attention mechanisms with soft physical constraints, enabling rapid inference with relatively low network complexity. The second module, Hybrid MLP-XGBoost Regressor module (HMXR), is established for aerodynamic coefficient prediction by decoupling aerodynamic response prediction from flow-field prediction. In this way, the prediction accuracy of aerodynamic coefficients is improved while detailed flow-field reconstruction capability is retained. The results show that ParaAero-Net performs effectively in small-sample aerodynamic prediction of two-dimensional airfoils and achieves good reconstruction of both flow-field structures and aerodynamic coefficients. For the cases considered in this study, a maximum fitting accuracy of 99.8% is achieved.