Abstract: Back propagation (BP) neural networks are established to predict the aerodynamic performance of an airfoil. Three typical types of design of experiments (DOE) methods, i.e., the factorial, orthogonal and uniform DOE, are employed to sample the training airfoils. The emphasis is laid upon the effect of the three DOE methods on the prediction accuracy of BP neural networks. The established BP neutral networks are finally applied to the optimization design of airfoil FX 63-167.The results show that, the factorial, orthogonal and uniform DOE have test-sample-averaged errors of 0.002%, 0.029% and 0.023% respectively in the case of small numbers of factors and levels. In the case of large numbers of factors and levels, the factorial DOE is unaccep-table due to excessive training points, while the orthogonal and uniform DOE have errors of 0.42% and 0.15%, respectively, indicating that the uniform DOE is the best method among the three, which can be coupled with the BP neural network for the prediction of the airfoil aerodynamic performance. The optimized airfoil shows a higher lift-drag ratio than the original one within the wide range of angles of attack from 0° to 18°. The lift-drag ratio increases by 4.38%, 1.38% and 5.51% respectively at angles of attack of 1°, 4° and 15°. The proposed method and conclusion can be extended to the multi-parameter optimization design of the turbo-machinery.

Key words: neural networks, training sample, design of experiments, airfoil aerodynamic performance, optimization design