Prediction of airfoil dynamic stall unsteady loads using dynamic neural networks architectures

Abstract This study investigates the application of dynamic neural networks for predicting unsteady aerodynamic loads on airfoils under dynamic stall conditions. The dynamic stall phenomenon is characterized by highly nonlinear and transient flow separation at high angles of attack. It poses significant challenges for traditional modeling approaches such as Computational Fluid Dynamics (CFD) and semi-empirical methods. To address these limitations, this work innovatively proposes a reduced-order model, trained exclusively on experimental data, that uses dynamic neural networks to rapidly and reliably evaluate aerodynamic loading on pitching airfoils under nonlinear dynamic stall conditions. Two neural network architectures, Time-Delay Neural Networks (TDNN) and Recurrent Neural Networks (RNN), are trained on experimental wind tunnel data for a NACA0012 airfoil undergoing pitching oscillations. The TDNN incorporates time-delayed inputs to capture temporal dependencies, while the RNN leverages internal feedback loops to model sequential aerodynamic behavior. Both models were trained using the Scaled Conjugate Gradient (SCG) algorithm and evaluated on their ability to predict lift, drag, and pitching moment coefficients. Results indicate that both architectures effectively capture dynamic stall’s nonlinear and hysteretic characteristics, with the RNN demonstrating marginally superior performance due to its inherent memory retention. The models achieved high accuracy in predicting lift and drag. However, challenges persisted in estimating the pitching moment, particularly in light stall regimes, likely due to noise sensitivity in moment data. Generalization tests on unseen cases confirmed the neural network models’ robustness in severe dynamic stall scenarios, including vortex shedding and flow reattachment.