Comparison of surrogate models for turbomachinery design

Abstract

This article addresses the issue of selecting surrogate models suitable for the global optimization of turbomachinery flows. As a first step towards this goal the analysis of a family of 2D flows on a two-parameter design space is presented. Four types of surrogate models are considered: least square polynomials, artificial neural networks (multi-layer perceptron and radial basis function) and Kriging. Discussed is the ability of these surrogate functions to give a satisfactory description of the exact function of interest on the design space, during a global optimization. The number of CFD evaluations for an adequate description of the exact function is presented