Predicting airfoil dynamic stall loads using neural networks

Dynamic stall is an aerodynamic regime characterized by loss of airfoil lift, drag increment, and abrupt changes in the pitching moment. Such effects can couple with structural dynamics where perturbations can be easily amplified, making this a critical phenomenon that jeopardizes operational safety. Hence, there is always the need to constantly study the basics of dynamic stall and provide newer predictive models that can take advantage of the current interest peak in artificial intelligence. The present work builds upon that need, exploring the ability of a simple feed-forward network to predict the oscillation cycle of a pitching airfoil experiencing from light to deep stall of a NACA0012 airfoil close to a Reynolds number of approximately 1.4 × 10 6 . The proposed neural network uses the angle of attack and its rate of change as inputs, then estimates the whole aerodynamic cycle at once, outputting an aggregated vector of drag, lift, and pitching moment coefficients. The training phase was conducted using a database containing several conditions obtained from experimental tests, with a strict convergence criterion of R 2 = 0.99 for both training and test datasets. Results show that the neural network, even in the least-performing conditions, can capture the aerodynamics and overall tendencies, even if some dynamics are underrepresented in the training dataset. The present work brings down the complexity of methodology while demonstrating that a simplistic architecture can still offer an accurate dynamic stall model. • A feedforward neural network can accurately predict the aerodynamics of a pitching airfoil in dynamic stall conditions. • The neural network model is trained using experimental data, showing its ability to generalize well under various conditions. • The FNN weights indicate an overdependence on the pitching rate and a clear learning segmentation in the output layer.