Sensitivity Analysis of Nested Multiphysics Models Using Polynomial Chaos Expansions

International audienceUncertainty quantification in computational mechanics has received much attention in the past few years. In this context, global sensitivity analysis (GSA) aims at quantifying which random input parameters (or combinations of parameters) are the most important in order to explain the variability of the model response. The present paper considers models of complex systems which are made of nested submodels, i.e. where input parameters of some submodels are obtained as the output of other submodels. A general framework is proposed in order to represent all the intermediate random variables (i.e. those variables that are obtained as response quantities of first-level submodels) by their polynomial chaos expansions. The coefficients of the various PC expansions are computed by regression and combined in order to obtain analytically the sensitivity (Sobol') indices. The crucial point is the representation of the joint PDF of intermediate input random variables, for which a Nataf transform is used. The approach is illustrated by sensivity analysis of a multiscale micromechanics-based model of concrete elastic properties