Response surface methodology for sensitivity and uncertainty analysis: performance and perspectives

Two main aspects have to be taken into account in studying a nuclear accident scenario when using nuclear safety codes as an information source. The first one concerns the behavior of the code response and the set of assumptions to be introduced for its modelling. The second one is connected with the uncertainty features of the code input, often modelled as a probability density function (pdf). The analyst can apply two well-defined approaches depending on whether he wants major emphasis put on either of the aspects. Response Surface Methodology uses polynomial and inverse polynomial models together with the theory of experimental design, expressly developed for the identification procedure. It constitutes a well-established body of techniques able to cover a wide spectrum of requirements, when the first aspect plays the crucial role in the definition of the objectives. Other techniques such as Latin hypercube sampling, stratified sampling or even random sampling can fit better, when the second aspect affects the reliability of the analysis. The ultimate goal for both approaches is the selection of the variable, i.e. the identification of the code input variables most effective on the output and the uncertainty propagation, i.e. the assessment of the pdf to be attributed to the code response. The main aim of this work is to present a sensitivity analysis method, already tested on a real case, sufficiently flexible to be applied in both approaches mentioned