Experimental design for the validation of kriging metamodels in computer experiments

Metamodels (or emulators) are statistical tools for the analysis of large complex simulation models. They consist of a Gaussian (or second order) process (kriging) fitted to a designed set of simulator runs. Once an emulator has been built, it is important that it is validated against some independent runs of the simulator. This paper considers the design of experiments for the validation of the fitted metamodel. All the proposed designs are based on maximin Latin hypercubes and add an extra criterion to be optimised based on the distances between the points in the validation and original designs. Simulation experiments are carried out to determine how well each design performs against the alternative criteria.