Incorporating Physical Models for Dynamic Stall Prediction Based on Machine Learning

Abstract

Unsteady aerodynamic prediction is an important part of modern aircraft safety and control law design. To improve the accuracy and efficiency for unsteady aerodynamic prediction of aircraft at high angles of attack, this paper proposed a machine learning framework based on multifidelity methods. The framework combines the linear dynamic derivative model and the fuzzy neural network model, which can achieve higher prediction accuracy under sparse experimental states. A series of wind-tunnel tests was carried out for the pitching motions of NASA Common Research Model at high angles of attack, to obtain steady and unsteady aerodynamic loads. These experimental data are used to verify the prediction accuracy of the unsteady model in a wide range of oscillation amplitude, frequencies, and mean angles of attack. The results show that the method has good generalization capability for the parameters of interest. At the same time, the comparison with the prediction results only from high-fidelity data shows that the proposed method can effectively reduce the amount of data required for the model of training and improve the modeling robustness to different types of motions.