Intuitive and feasible geometric representation of airfoil using variational autoencoder

Abstract Airfoil shape optimization is crucial for improving aerodynamic performance in advanced aircraft designs. Given the extensive functional evaluations required for optimization, surrogate modeling is widely used to alleviate computational burden. However, greater flexibility in airfoil parameterization often requires a larger number of design variables, leading to the challenge known as the curse of dimensionality in surrogate modeling. In recent years, generative models such as generative adversarial networks and variational autoencoders have shown potential to represent large design spaces with compact design variables. However, these models still exhibit limited feasibility and intuitiveness due to their high model capacity, which in turn degrades the efficiency of design optimization. To address this issue, we have developed a novel airfoil parameterization method using a variational autoencoder. The proposed method improves feasibility by using architecture modeling to separate the generation of thickness and camber distributions, resulting in smooth and nonintersecting airfoils. It also improves intuitiveness by using a physics loss function that aligns latent dimensions with geometric features of the airfoils. Notably, extensive comparative analyses validate the effectiveness of our method in terms of flexibility, parsimony, feasibility, and intuitiveness, leading to increased efficiency in aerodynamic design optimization.