A Comparison of Surrogate Models in the Framework of an MDO Tool for Wing Design

The replacement of the analysis portion of an optimization problem by its equivalent metamodel usually results in a lower computational cost. In this paper, three dierent metamodels are compared against the conventional non-approximative approach: quadratic interpolation based response surfaces, Kriging and Articial Neural Networks (ANN). The results obtained from the solution of three dierent case studies based on aircraft design problems reinforces the idea that quadratic interpolation is only well suited to very simple problems. At higher dimensionality, the usage of the more the complex Kriging and ANN models may result in considerable performance benets. Nomenclature b=2 Wing semispan, m c;ci Coecients for polynomial interpolation cbs Wing breakstation chord, m croot Wing root chord, m ctip Wing tip chord, m f (x) Regression model (Kriging) g (x) Constraint function nDV Number of design variables ns Number of samples nt Number of terms in polynomial interpolation/regression approximation qk(x) Values of regression functions at sample locations (Kriging) R (w; x; ) Correlation model (Kriging) sk Vector of independent variable samples (Kriging)