Sensitivity analysis with dependence and variance-based measures for spatio-temporal numerical simulators

Matthias de Lozzo (1) , Amandine Marrel (1)

1 DER - Département Etude des Réacteurs

In a case of radioactive release in the environment, modeling the radionuclide atmospheric dispersion is particularly useful for emergency response procedures and risk assessment. For this, the CEA has developed a numerical simulator, called Ceres-Mithra, to predict spatial maps of radionuclide concentrations at different instants. This computer code depends on many uncertain scalar and temporal parameters, describing the radionuclide, release or weather characteristics. The purpose is to detect the input parameters the uncertainties of which highly affect the predicted concentrations and to quantify their influences. To this end, we present various measures for the sensitivity analysis of a spatial model. Some of them lead to as many analyses as spatial locations (site sensitivity indices) while others consider a single one, with respect to the whole spatial domain (block sensitivity indices). For both categories, variance-based and dependence measures are considered, based on recent literature. All of these sensitivity measures are applied to the CM computer code and compared to each other, showing the complementarity of block and site sensitivity analyses. Finally, a sensitivity analysis summarizing the input uncertainty contribution over the entirety of the spatio-temporal domain is proposed.

Mots clés: Spatio-temporal models; Global sensitivity analysis; Sobol' indices; Dependence measures; Site & block sensitivity measures; Gaussian process model

Domaines: Applications [stat.AP]; Mathématiques [math]; Statistiques [stat]; Méthodologie [stat.ME]

Matthias de Lozzo, Amandine Marrel. Sensitivity analysis with dependence and variance-based measures for spatio-temporal numerical simulators. Stochastic Environmental Research and Risk Assessment, 2017, 31, pp.1437-1453. ⟨hal-01253686⟩