Fast Airfoil Design Based On Multi-output Gaussian Process Regression

Airfoil design is an important step in the aircraft design, affects the flying qualities of the aircraft.Wind tunnel experiments and CFD technology are the traditional airfoil design methods. However,they are usually costly and time-consuming. In order to reduce costs and improve the airfoil designefficiency, this thesis applies multi-output Gaussian process (MOGP) regression in the area ofmachine learning to the airfoil aerodynamic performance fast evaluation and airfoil fast design, as asupplement to the existing airfoil design method. MOGP uses the convolution of Gaussian baseprocesses and smooth kernel functions to model each output, and therefore captures the correlationbetween the multiple outputs by sharing some common base processes.In this thesis, a group of NACA series airfoils is selected, and the corresponding lift coefficients,drag coefficients and surface pressure distributions are calculated. Based on these datas, we conductedthe experiments of airfoil aerodynamic performance evaluation and airfoil fast design. The experimentof airfoil aerodynamic performance evaluation uses the airfoil shape parameters as input datas topredict airfoil lift coefficients and drag coefficients; airfoil fast design uses the pressure distributionsof airfoil surface as input datas to predict airfoil shape parameters. We have established the regressionin models for MOGP, single-output Gaussian process (SOGP), BP and RBF artificial neural networkto model the non-linear relationship between airfoil shapes and aerodynamic performance, andcompared their prediction accuracy. The experimental results show that MOGP has better predictionaccuracy compared with the other alternatives when the multiple outputs have significant correlationin the airfoil aerodynamic performance evaluation. In the airfoil fast design, MOGP has the similardesign accuracy to SOGP, but outperforms the BP and RBF networks when the multiple outputs havelittle correlation. For all these reasons, the correlation between multiple outputs should be consideredin airfoil design when significant correlation is observed across outputs, in order to improvingprediction accuracy.