Inverse Design of Airfoil Using a Deep Convolutional Neural Network

Abstract

This paper proposes an approach to perform the inverse design of airfoils using deep convolutional neural networks (CNNs). The conventional approaches are based on the solution of differential equations, which are either difficult to solve or take a tedious procedure to obtain the solution. As shown in this paper, the deep learning technique can be used effectively to obtain the airfoil shape from the coefficient of pressure distribution. More specifically, the CNN is applied due to its ability to handle any airfoil geometry without the need for complex parametrization. In the training phase, the pressure coefficient distribution is fed as input to the CNN to obtain a prediction model for the airfoil shape. In the testing phase, a new pressure coefficient distribution is given to the CNN model, generating an airfoil shape that is very close to the associated airfoil with an error of less than 2% for most of the cases. An extensive investigation of the effects of various hyperparameters in the CNN architectures is performed. The results show that the CNN method is very efficient in terms of computational time and shows a competitive prediction accuracy. It thus constitutes an attractive new method in the inverse design of airfoils, especially when existing data can be readily exploited.