Surrogate Modeling of Stochastic Functions - Application to computational Electromagnetic Dosimetry

Abstract

Metamodeling of complex numerical systems has recently attracted the interest of the mathematical programming community. Despite the progress in high performance computing, simulations remain costly, as a matter of fact, the assessment of the exposure to radio frequency electromagnetic fields is computationally prohibitive since one simulation can require hours. Moreover, in many engineering problems, carrying out deterministic numerical operations without considering uncertainties can lead to unreliable designs. In this paper we focus on the surrogate modeling of a particular type of computational models called stochastic simulators. In contrast to deterministic simulators which yield a unique output for each set of input parameters, stochastic simulators inherently contain some sources of randomness and the output at a given point is a probability density function. Characterizing the stochastic simulators is even more time consuming. This paper represents stochastic simulators as a stochastic process and describes a metamodeling approach based on the Karhunen-Lo\`eve spectral decomposition.