Joint metamodeling for sensitivity analysis of continuous and stochastic inputs of a computer code

Abstract

To perform the global sensitivity analysis of a complex and cpu time expensive code, a mathematical function built from a small number of simulations referred to as a metamodel can be used to approximate the code. In some applications like oil reservoir simulations, the code output can depend on complex stochastic inputs such as random permeability fields. This paper proposes a new metamodeling approach to perform global sensitivity analysis of both scalar and such type of stochastic input. A joint metamodeling based on two Gaussian process metamodels is proposed to model the mean and the variance. Then, the sensitivity indices of scalar and stochastic inputs are estimated from this joint metamodeling. An application on a reservoir simulator illustrates the overall methodology.