Scaling properties of multi-fidelity shape optimization algorithms

Abstract

Multi-fidelity optimization can be utilized for efficient design of airfoil shapes. In this paper, we investigate the scaling properties of algorithms exploiting this methodology. In particular, we study the relationship between the computational cost and the size of the design space. We focus on a specific optimization technique where, in order to reduce the design cost, the accurate high-fidelity airfoil model is replaced by a cheap surrogate constructed from a low-fidelity model and the shape-preserving response prediction technique. In this study, we consider the design of transonic airfoils and use the compressible Euler equations in the high-fidelity computational fluid dynamic (CFD) model. The low-fidelity CFD model is same as the high-fidelity one, but with coarser mesh resolution and reduced level of solver converge. The number of design variables varies from 3 to 11 by using NACA 4-digit airfoil shapes as well as airfoils constructed by Bézier curves. The results of the three optimization studies show that total cost increases from about 12 equivalent high-fidelity model evaluations to 34. The number of high-fidelity evaluations increases from 4 to 9, whereas the number of low-fidelity evaluations increases more rapidly, from 600 to 2000. This indicates that, while the overall optimization cost scales more or less linearly with the dimensionality of the design space, further cost reduction can be obtained through more efficient optimization of the surrogate model.