Data-driven inverse aerodynamic design of blade profiles via conditional generative adversarial networks

In the field of turbomachinery engineering design, profile design is a complex process highly dependent on iterative optimization, often requiring the experience of human designers to repeatedly adjust the profile shape to meet requirements. Aiming at the need to directly generate shapes that satisfy target specifications in practice, the conditional Generative Adversarial Network (cGAN) demonstrates unique advantages by its data-driven direct design synthesis capability. This paper proposes a data-driven inverse aerodynamic design method for blade profiles based on cGAN and Multilayer Perceptron (MLP). In this method, MLP is utilized to learn the mapping from design parameters to performance parameters. By associating profile design parameters with precomputed aerodynamic performance, a cGAN is employed to learn to generate geometries that meet the specified aerodynamic performance. MLP is also used to provide the dataset for training the cGAN and to test the generation capability of the cGAN. Subsonic design validation confirms that the trained cGAN generator can directly produce profile geometries meeting the specified performance target, eliminating complex iterations and significantly reducing design time, thereby substantially boosting design efficiency.