Comparing sampling strategies for aerodynamic Kriging surrogate models

Abstract

AbstractIn aerodynamic applications often evaluations of an expensive computer simulation like a CFD solver are needed for a whole range of input parameters. Dense computations to describe the global behavior of an objective function are out of reach due to limited computational resources. Surrogate models like the Kriging method allow an interpolation of collected data and a global approximation. Adaptive sampling strategies can reduce the number of required samples for accurate and efficient surrogate models by automatically identifying critical or too coarse sampled regions of the input domain. We compare different existing sampling strategies as well as new theoretical methods using a dense set of validation data in order to gain a deeper understanding of optimal sample distributions and lower error boundaries.