Generation and optimization of aerodynamic shapes for aircraft based on generative adversarial networks

Abstract

To address the critical issues of low efficiency and lack of innovation in traditional aircraft aerodynamic shape design, this paper proposes an intelligent design method based on Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and TensorFlow 2.x. A foundational dataset containing 12,000 airfoil samples is constructed and augmented to 24,000 samples covering transonic, supersonic, unmanned aerial vehicle (UAV), and airliner configurations. The generator employs a series of fully connected layers with batch normalization and LeakyReLU activation functions. Experimental results show that the model converges after 150 training epochs, with the generator loss stabilizing between 3.8 and 4.1 and the discriminator loss between −5.5 and −4.5. The optimized airfoil achieves a stable lift-to-drag ratio of 107.8, representing a 16.8% improvement compared with the baseline NACA 0012 airfoil. Gradient variation curves, parameter update heatmaps, geometric feature distributions, and manufacturability assessments are introduced to enhance model transparency and engineering practicality. Compared with state-of-the-art methods, including VAEGAN and conditional WGAN-GP, this method demonstrates superior multi-condition generalization, training stability, and interpretability, providing a reliable technical pathway for rapid aircraft aerodynamic shape design.