Convolutional Variational Auto Encoder Based Deep Neural Network for Calculation of Airfoil Aerodynamic Parameters

Abstract

The Convolutional Neural Network (CNN) method is capable of image processing and is widely used today for aerodynamic meta-modelling tasks. CNN can predict the aerodynamic properties of an airfoil based on a large enough dataset. Variational Auto Encoder (VAE), which belongs to the family of Generative Adversal Networks (GANs), also has the capability of image processing and can be used for extracting the best features on the basis of two additional layers namely, Mean and Variance. In this thesis, convolutional variational auto encoder is applied to the UIUC aerodynamic dataset to predict the lift-to-drag ratio of airfoils with various angles of attack. This work also employs the Convolutional Neural Network method to calculate the lift-to-drag ratio on similar dataset. The efficiency and accuracy of these two methods are compared and discussed in this thesis. It is demonstrated that the VAE method can maintain a relatively competitive level of accuracy while being far more time efficient (1.26 times in training and 1.76 times in testing on CPU while 1.22 times in training and 1.38 in testing on GPU) both in terms of training and testing than CNN method.