Deep Generative and Probabilistic Surrogate Modeling for Uncertainty-Based Airfoil Design Optimization

Abstract

Uncertainty-based aerodynamic shape optimization is challenged by the curse of dimensionality and high computational costs. To address this, we propose a deep generative and probabilistic surrogate modeling for optimizing subsonic airfoils under uncertainty. A variational autoencoder learns a compact latent representation of airfoil geometries for dimensionality reduction and design generation, while a mixture density network predicts aerodynamic uncertainties under varying flight conditions. The proposed framework achieves a mean drag prediction [Formula: see text] of 0.999 and a standard deviation prediction [Formula: see text] of 0.970. Optimization results show that deterministic, robust, and reliability-based designs closely match those obtained from Monte Carlo simulations with 5000 samples, while our method requires only 200 samples per design, trained over the global design space. This shows the potential to mitigate dimensionality and reduce computational costs in uncertainty-based design optimization.