Adaptive generative design optimization method for compressor airfoils

The Bézier Generative Adversarial Network, a dimensionality reduction technique combining Generative Adversarial Networks with a Bézier parametric curve formulation, achieves both low-dimensional characterization and high-fidelity smoothness in airfoil representation. However, the latent feature vectors resulting from this geometric reduction lack physical interpretability, and their variable boundaries are arbitrary defined, which limits further gains in optimization efficiency. To address this issue of arbitrary post-reduction boundary setting, this paper proposes an enhanced adaptive Bézier-GAN method. The proposed approach integrates an adaptive variable-boundary expansion strategy with a tabu search technique. By dynamically adjusting the sampling boundaries and guiding the optimization direction, it effectively resolves the arbitrariness inherent in setting variable boundaries after dimensionality reduction. When applied to compressor airfoil optimization, the improved method reduces the minimum total pressure loss coefficient by 39.1% and 13.8% relative to conventional dimensionality reduction methods and the standard Bézier-GAN, respectively. This demonstrates its superior optimization capability.