Fast Flapping Aerodynamics Prediction using a Recurrent Neural Network

Being able to calculate aerodynamic coefficients of oscillating airfoils is a prerequisite for the design of rotor blades and bio-inspired micro air vehicles. Numerous methods have been used throughout the years, however, the intricacy of the underlying physics that govern the flapping motion makes many of these models computationally expensive, resulting in a need for faster and less intensive approaches. With this objective in mind, the present work proposes an implementation of a recurrent neural network to obtain the lift and moment coefficients of an airfoil undergoing flapping motion. There has been a surge in the use of neural networks in the field of aerodynamics for their capability of, once trained, producing fast reasonably accurate predictions. The proposed network is fed with the flow's Reynolds number, reduced frequency, and nondimensional amplitude, as well as the time history of the angle of attack. From this, the network can predict the evolution of the mentioned aerodynamic coefficients with time. The neural network is trained using data from a panel code, which can be used when no massive flow separation is present. Despite its limitations, the panel method is used, since it allows for the creation of an extensive training dataset promptly. Results show that the proposed neural network is indeed capable of predicting the aerodynamic coefficients with sufficient accuracy, thus reinforcing the potential artificial intelligence-based solutions have for quick aerodynamic computation. Future work looks at retraining the network using experimental data and identifying if the inputs are sufficient to capture real-fluid physics. Other improvements to the network are on the horizon, such as taking the airfoil geometry as input and increasing the network's flexibility.