A comparison study of two multifidelity methods for aerodynamic optimization

Abstract

Industrial aerodynamic design applications require multiobjective optimization tools able to provide design feedback to the engineers. This is true especially when optimization studies are carried out during the conceptual design stage. The need for fast optimization methods has led to the development of multifidelity methods in a surrogate based optimization environment. Multifidelity tools have the potential to accelerate the design process, primarily due to the lower cost associated with the low fidelity tool. In addition to this, the design stage is shortened as mature and reliable high fidelity design information is provided earlier in the design cycle. Despite this high potential of these methods, there is no explicit comparison available in the literature between multifidelity surrogate based optimization tools for industrial aerodynamic problems. This paper aims at providing a direct comparison between two multiobjective multifidelity surrogate based optimization methods developed by our group. The first approach uses a trust region formulation for efficient multiobjective that does not require gradients. The second is using the concept of expected improvement to perform fast design space exploration based on a novel Kriging modification for multifidelity data. The tools are applied in two aerodynamic design problems: optimization of a high lift configuration in respect to maximum lift maximization and an airfoil design for transonic cruising conditions. These problems feature characteristics of industrial interest. They involve difficult physical analyses in the case of the high lift configuration and a more complex optimization formulation due to the increased dimensionality in the case of the transonic airfoil. Our presented methods are compared against a CFD-based optimization, a surrogate based optimization using only high fidelity data and a multifidelity surrogate based optimization based on Co-Kriging. Early results suggest that the trust region method can quickly provide improved designs leading to an efficient Pareto front. The expected improvement based method shows fast exploration attributes and a wide Pareto front.