An estimation algorithm for fast kriging surrogates of computer models with unstructured multiple outputs

Abstract

Computationally intensive computer models are used in many areas of engineering. In order to speed up the investigations, fast statistical surrogates have been developed in the literature. The surrogates addressed in this paper incorporate a general and unstructured covariance, best suited for modeling nonlinear and nonstationary multiple outputs. We propose an efficient algorithm to cope with the estimation of a large number of parameters. Then multivariate kriging is used to construct the fast surrogate. This algorithm can be embedded in both maximum likelihood and cross-validation estimation methods. We compare the proposed method with a current method based on principal components. The methodology is illustrated with a mechanical engineering application involving a vehicle suspension system.