Numerical study of algorithms for metamodel construction and validation

In some studies requiring predictive numerical models, it can be advantageous to replace cpu time expensive computer models by cpu-inexpensive mathematical functions, called metamodels. In this paper, we focus on the Gaussian process metamodel whose construction requires an initial design of computer model simulations. A numerical study compares different types of Latin hypercubes maximized relatively of metamodel predictivity. The metamodel validation step consists in estimating the true metamodel predictivity with a minimum number of additional calculations. In this goal, we propose and test an algorithm which optimizes the distance between the test points and points of the initial training base.