A novel deep-learning-based pressure distribution prediction approach of airfoils

Abstract

Pressure distribution is a crucial flow characteristic and a key consideration in supercritical airfoil design. Traditionally, obtaining the pressure distribution involves time-consuming and computationally expensive wind tunnel experiments and computational fluid dynamics calculations. This study proposes a deep-learning-based approach to directly map input geometric information to the pressure distribution output, thereby avoiding costly wind tunnel experiments and iterative computational fluid dynamics simulations based on Navier–Stokes equations to address these challenges. Conventional surrogate models typically focus on predicting simple force factors, such as lift and drag coefficients, or require the conversion of airfoil data into images for model training. The novel approach utilizes a Variational Autoencoder for pressure distribution characteristic extraction and reconstruction from feature variables. Unlike conventional models, this approach avoids image conversion and employs a radial basis function neural network for effective mapping. The model exhibits good fitting and generalization capabilities on both training and test datasets, offering a promising solution for rapid pressure distribution prediction in airfoil design. This novel deep-learning-based approach advances airfoil design methodologies, offering significant advantages in computational efficiency and performance prediction. By directly mapping geometric information to pressure distribution, it provides an innovative and promising tool for airfoil design optimization.