Development of a Conditional Generative Adversarial Network for Airfoil Shape Optimization

In the field of aerodynamics, shape optimization is a very critical step. However, Computational Fluid Dynamics (CFD) simulation tools are computationally expensive to be directly employed in the optimization process. From this observation, the current work aims to leverage novel approaches to achieve accurate shape optimization for airfoil designs in a less computationally demanding manner. In particular, the recent advances in generative adversarial network (GAN) have been shown to enable efficient design space exploration by generating realistic samples using only a small set of examples without the explicit dictation of a large amount of design variables. In this work, the original GAN formulation is being elevated from a simple generative tool to a powerful semi-supervised learning framework. Specifically, a conditional generative adversarial network (CGAN) is adapted for the airfoil shape optimization task following the airfoil shape generation task. In contrast to the original GAN, CGAN uses labeled data for its training and generates realistic samples for any particular class that exists in a given data set. By labeling each airfoil shape with precalculated aerodynamic characteristics such as lift-to-drag ratio (L/D) and structural requirements (shape area), it is possible to guide the shape generation process toward particular classes of labeled samples, thereby providing a novel approach to shape optimization.