FoilGen2: Learning Coupled Latent Spaces for Hybrid and Performance-Driven Airfoil Generation

Airfoil design plays a key role in aerospace and renewable energy applications. However, it remains bottlenecked by slow, iterative trial-and-error evaluation using CFD or experiments, which limits rapid exploration of the design space. FoilGen2 learns separate latent spaces for geometry and performance, linked via a neural mapper, enabling hybrid design, smooth interpolation, and performance-driven synthesis. Experiments on 6,359 airfoils demonstrate geometric reconstruction errors below 0.8% of chord, lift coefficient errors under 3%, drag coefficient errors below 0.3%, and L/D errors under 3%. These results highlight the effectiveness of coupled latent spaces for flexible, aerodynamically coherent airfoil generation.