AI-driven prediction of aerodynamic coefficients using VAE-GAN and MLP models for 2D airfoils

The key highlights of our study include: • A VAE-GAN framework generates novel, smooth, and simulation-ready 2D airfoils. • An MLP surrogate predicts lift and drag with high accuracy in milliseconds. • GAN-augmented training improves generalization to unseen airfoil geometries. • he workflow reduces aerodynamic evaluation time by orders of magnitude vs CFD. Accurate prediction of aerodynamic coefficients is essential for airfoil design, yet high-fidelity Computational Fluid Dynamics (CFD) simulations are computationally expensive and unsuitable for real-time or large-scale screening. This study presents a novel end-to-end AI-driven framework that integrates generative modeling, reduced-order aerodynamic simulation, and supervised learning to enable fast and generalizable aerodynamic prediction across diverse airfoil geometries. A variational autoencoder–generative adversarial network (VAE-GAN) was trained on UIUC and NACA 4-digit airfoil databases to generate smooth, physically plausible airfoils, extending the design space beyond conventional parametric families. After geometric and aerodynamic filtering, 658 synthetic airfoils were retained and combined with real profiles, yielding approximately 87,000 samples evaluated using XFOIL, assuming two-dimensional, steady, incompressible flow, at three Reynolds numbers (3 × 10⁵, 5 × 10⁵, 8 × 10⁵) and 21 angles of attack (−5° to 15°). A Multi-Layer Perceptron (MLP) surrogate was trained to predict lift (CL) and drag (CD) directly from full airfoil geometry, angle of attack, and Reynolds number (404 input features), comparing with XFOIL, achieving RMSE values of 0.1817 for CL and 0.0483 for CD, with R² = 0.9966 and R² = 0.9829, respectively. To assess physical fidelity beyond reduced-order modeling, selected cases were validated using steady two-dimensional RANS simulations performed with ANSYS Fluent, based on the finite-volume method and the SST k–ω turbulence model. Generalization was evaluated using the NACA0012 airfoil at Re = 5 × 10⁵, excluded from training, yielding R² values of 0.9887 for CL and 0.8261 for CD relative to CFD. Inference times were below one second per sample on a standard CPU. The results demonstrate that combining generative geometry augmentation with supervised surrogate modeling enables accurate, physically consistent, and computationally efficient aerodynamic prediction, supporting rapid design iteration and real-time multidisciplinary applications. To create your abstract, type over the instructions in the template box below. Fonts or abstract dimensions should not be changed or altered.