Quantification and reduction of uncertainty in aerodynamic performance of GAN-generated airfoil shapes using MC dropouts

Abstract

Generative adversarial network (GAN) models are widely used in mechanical designs. The aim in the airfoil shape design is to obtain shapes that exhibits the required aerodynamic performance, and conditional GAN is used for that aim. However, the output of GAN contains uncertainties. Additionally, the uncertainties of labels have not been quantified. This paper proposes an uncertainty quantification method to estimate the uncertainty of labels using Monte Carlo dropout. In addition, an uncertainty reduction method is proposed based on imbalanced training. The proposed method was evaluated for the airfoil generation task. The results indicated that the uncertainty was appropriately quantified and successfully reduced.