Fast Flapping Aerodynamics Prediction Using a Recurrent Neural Network

Abstract

One of the major tasks of aerodynamics is the study of the flow around airfoils. While most conventional methods deal well with steady flows, unsteady airfoils, like the ones on helicopter blades, are subject to such complex dynamic flows that their study can impose substantial difficulties. However, recent applications of machine learning, in the form of neural networks, have shown very promising results when dealing with complex dynamic aerodynamic phenomena. For this reason, this paper proposes the implementation of a recurrent neural network for the time-wise prediction of the lift, momentum, and drag coefficients for an airfoil subject to plunging motion, using the Re, k, h, kh and the time history of the effective angle of attack as inputs. Results from early training already suggest the network’s capability to approximate the desired outputs, even if with some limitations. However, the network configuration is flexible enough to be fed with either experimental or numerical data in the future.

Keywords

• unsteady aerodynamics
• recurrent neural network
• potential flow