Surrogate models based on function and derivative values for aerodynamic global optimization

Abstract : The application of global algorithms in aerodynamic shape optimization is still today very limited due to the required CPU time. To avoid this limitation, one method is to replace the objective function by a surrogate model. One important achievement of the recent years is the development of accurate and robust surrogate models, minimizing the quantity of information (computed via costly CFD simulations) required for their construction. In this article, we propose to use derivatives in addition to function values to determine the meta model. With adjoint methods, these derivatives may be obtained at a reduced cost, independent of the number of design variables. Three distinct meta model constructions including derivatives are presented and adapted. Then some aerodynamic design applications are presented. The benefit of including derivatives in the surrogate models is clearly evident.