This paper describes an efficient neural network model called AirfoilNet, for enhancing its predictivity of the aerodynamic parameters (e.g., the lift and drag coefficients) of airfoils. To address these challenges such as limited aerodynamics datasets and end-to-end training and inference, a diffusion model, Denoising Diffusion Aerodynamic Model (DDAM), is designed to generate additional synthetic datasets to support the application of deep learning methods to improve the prediction of the aerodynamic parameters. AirfoilNet first encodes grayscale airfoil images into low-dimensional representations, incorporating the relevant parameters, including Reynolds number, angle of attack, and geometric coordinates of airfoils, for computations. Then, the extracted airfoil features are fed into the prediction modules for a complete end-to-end prediction of the aerodynamic parameters. To further improve the accuracy and flexibility of the predictions, AirfoilNet incorporates two different prediction modules, namely the Gated Recurrent Unit Network and the Residual Multilayer Perceptron, for iterative optimization and refined prediction schemes. It is demonstrated in experiments that AirfoilNet has advantages in inference speed, generalization ability, and accuracy for the prediction of the aerodynamic parameters. Furthermore, the synthetic data generated by DDAM effectively improve the overall accuracy and robustness of AirfoilNet.
Cai et al. (2025) studied this question.