Analysis of Generative Adversarial Networks for Data Driven Inverse Airfoil Design

EasyChair Preprint no. 5975

Abstract

Data-Driven Methods have led to new approaches in the field of Aerodynamic Design and have recently found success in Inverse Design applications. The conventional Inverse Design methods are analytically complex and mathematically demanding to formulate. This research attempts to perform Inverse Airfoil Design using Generative Adversarial Networks(GANs) with the objective to generate airfoil shapes that produce desired Pressure Distribution at given flow conditions. The Convolutional Neural Network-based Generator extracts features from the pressure coefficient profiles and predicts the corresponding airfoil shape coordinates. These deep ConvNet structures eliminate the problems posed by shape parameterization in classical methods and extract patterns from the data with finer details. This work examines the performance of three advanced Generative Adversarial Network architectures to obtain a model which is stable, computationally efficient and has competitive prediction accuracy. The candidate GANs include Wasserstein GAN, Boundary Seeking GAN and Bidirectional GAN. The networks are trained on a database of airfoil shapes and pressure coefficient distribution. It is shown that Boundary Seeking Generative Adversarial Network produces highly accurate results and is computationally least expensive to train.

Keyphrases: Generative Adversarial Networks, Inverse design, machine learning

@Booklet{EasyChair:5975,
  author = {Priyam Gupta and Prince Tyagi and Raj Kumar Singh},
  title = {Analysis of Generative Adversarial Networks for Data Driven Inverse Airfoil Design},
  howpublished = {EasyChair Preprint no. 5975},

  year = {EasyChair, 2021}}