Physics-informed neural networks with boundary-condition constraints and wake-region weighting for airfoil flow prediction

Conventional numerical simulation methods face significant limitations in efficiency when dealing with multiple-angle-of-attack airfoil simulations. To overcome these challenges, this research presents a parametric physics-informed neural networks (PINN) framework for predicting flow fields across multiple angles of attack. By encoding the angle of attack as an input parameter, the model learns a parametric surrogate of computational fluid dynamics (CFD)-generated mean-flow fields across multiple angles of attack, enabling continuous interpolation and extrapolation in the parameter space. To improve training stability and flow consistency, a physics-informed spatial attention strategy with wake-region weighting is developed to prioritize high-gradient wake regions, mitigating the “gradient pathology” prevalent in standard PINNs. Furthermore, a velocity-component weighting scheme is also implemented to address imbalanced convergence between velocity components. Comparative validation against CFD data demonstrates that the proposed framework achieves robust predictive accuracy for attached and mildly separated flows, with good agreement in pressure coefficient distributions and velocity field. The results demonstrate good predictive accuracy and training stability for attached and mildly separated mean flows, establishing a computationally efficient paradigm for rapid airfoil aerodynamic assessment.