Airfoil Flowfield Prediction Based on Proper Orthogonal Decomposition with Deep Learning

In aerodynamics, the flowfield around an airfoil is typically solved using experiments or computational fluid dynamics (CFD), while they are often computationally expensive. Reduced-order models (ROMs) can effectively balance the accuracy and efficiency of CFD, and deep learning can also be integrated with ROMs to approximate flowfields, thereby avoiding the need to solve the complex Navier–Stokes (N-S) equations. Using the NACA0012 airfoil as a case study, a proper orthogonal decomposition–gated recurrent unit (POD-GRU) model is proposed for the flowfield approximation purpose. For the approximation, the CFD simulation data is used for training, POD is employed for dimensionality reduction, and the GRU network is used to predict the coefficients for flowfield reconstruction. This is the first time that the proposed model has been used to predict the flowfield of an airfoil. The model achieves a prediction error within 8% of high-fidelity models, while its computational cost for a single operating condition is merely 0.4% of that required by traditional CFD simulations. This remarkable efficiency and accuracy make the method particularly suitable for applications requiring rapid response, such as airfoil control to mitigate airfoil stall, flow interference, and enhancing lift efficiency during flight. Furthermore, the model demonstrates robust predictive capabilities across varying inflow velocities and angles of attack, showing its significant potential for engineering applications where both speed and precision are critical.