Aerodynamic Design Optimisation Through Machine Learning—GCAG-VAEGAN

This paper presents, as a proof-of-concept, a novel geometrically constrained, aerodynamics-guided variational autoencoder generative adversarial network (GCAG-VAEGAN) framework for aerofoil design, aimed at improving the aerodynamic feasibility and numerical stability of generative aerofoils. This architecture enabled the model to incorporate aerodynamic awareness into the generative process, resulting in substantial improvements in numerical convergence and geometric quality. The proposed model combined geometry-constrained latent learning with an embedded aerodynamic feedback loop that guided the generation process. A discriminator module enforced geometric realism, while the physics predictor provided aerodynamic supervision. Together, these components were able to consistently generate smooth, geometrically stable aerofoils. Illustrative evaluation against traditional generative models and NACA aerofoils demonstrated clear, quantitative improvements in geometric validity, convergence behaviour, and a reduction in extreme geometries. The GCAG-VAEGAN achieved higher XFOIL convergence rates than both baseline GC-VAEGAN and conventional NACA aerofoils. These improved rates were accompanied by the generation of novel geometries comparable with high-performance aerofoils, such as the NACA 4412, despite limited initial aerodynamic conditions. Overall, the findings demonstrate the effectiveness of the integrated aerodynamic loop within the generation process, not only improving shape feasibility but also highlighting the potential for accelerating conceptual aerodynamic design exploration.