Abstract
A semi-intrusive approach to the surrogate modeling in aerodynamic shape optimization problems is presented and its approximating accuracy is analyzed quantitatively The method is semi-intrusive as it uses flow field information m the training of the surrogate models A set of CFD flow field samples obtained with design of experiment are processed through a proper orthogonal decomposition, a decomposition into orthonormal basis of the flow field covariance matrix, which recovers the sample data as the subspace spanned by the orthonomal basis The surrogate model is constructed by establishing a functional relationship between the set of design parameters which generated the sample set and the coefficients of the POD The radial basis function network was used for this purpose The resulting POD-RBFN surrogate model was tested for its approximating accuracy by two validation methods, leave-one-out cross-validation and validation with a fixed set of validating samples The results shows the POD-RBFN approach performs much better than a simple RBFN approach which dmodels the relationship between the design parameters and the scalar objective functions directly
