INTELLIGENT RECONSTRUCTION METHOD OF AIRFOIL FLOW FIELD BASED ON DEEP ATTENTION NETWORK

The computation of aerodynamic parameters using Navier-Stokes (NS) equations is notably time-consuming. To address this, a data-driven Deep Attention Network (DAN) is introduced for rapid reconstruction of steady flow fields over various airfoil shapes. To effectively represent geometric information of different airfoils, the airfoil profile grayscale images fed into the network are segmented into distinct patches, and embedding corresponding positional information. Subsequently, these embedded geometric vectors are processed through a Transformer encoder to extract attention-based geometric features specific to each airfoil. Finally, the extracted geometric features from the Transformer encoder, combined with flow coordinates and wall distance, are fused and input into a multi-layer perceptron to predict the velocity and pressure fields of the airfoil. Through quantitative and qualitative analysis of extensive experimental results, it is observed that the proposed deep attention network model possesses certain geometric interpretability. Furthermore, it showcases robust generalization capabilities and high prediction accuracy across a wide array of airfoil flow fields.