Deep learning–driven surrogate modeling framework for aerodynamic drag optimization

This study proposes an aerodynamic drag optimization scheme based on deep learning and surrogate modeling, aiming to achieve efficient shape design optimization without relying on high-fidelity CFD simulations or experimental data. Instead, synthetic data and machine learning methods are employed. Traditional aerodynamic design optimization methods typically require substantial computational resources and experimental data. In contrast, this research introduces autoencoders (AEs) and surrogate models (e.g., Gaussian Process Regression, GPR) for dimensionality reduction and modeling of shapes, which are then combined with optimization algorithms to enable efficient search in a low-dimensional latent space. Specifically, synthetic data are first generated to represent the relationship between shapes and drag; then, the autoencoder is used to reduce dimensionality, compressing high-dimensional shape information into a latent space. The surrogate model is subsequently trained to construct the drag response surface in the latent space, followed by Bayesian optimization and gradient-based optimization to identify the optimal shape. Experimental results demonstrate that the proposed method significantly reduces aerodynamic drag, with optimized shapes achieving approximately 20% reduction in drag coefficient while conforming to fundamental aerodynamic principles. Furthermore, the method exhibits flexibility and generality, providing effective optimization schemes for complex shapes in scenarios where CFD simulations or experimental data are unavailable. This study offers new perspectives for applying deep learning to aerodynamic design and provides strong support for advancing digital and efficient design processes.