Application of Sensitivity Analysis Techniques to a Low Dimensional Problem in the Frame of Severe Accidents

Abstract The quantification of uncertainties in Severe Accident (SA) simulations is crucial for ensuring the robustness and enhancing the reliability of nuclear reactor safety assessments. However, traditional UQ methods often lack the ability to discern the individual contributions of various input parameters to the overall uncertainty in simulation results. To this end, a sensitivity analysis complementary to the uncertainty one seems indispensable. In this framework, the present paper presents a twofold approach to address this challenge: on one side, by applying the so called single variable approach, in which the relationship between the selected output response and each input parameter taken individually is studied; on the other side, by integrating supervised Machine Learning (ML) techniques into the sensitivity analysis of SA simulations. The selected sensitivity analysis techniques are applied to the outcomes of SA simulations conducted as part of the recently concluded “Management and Uncertainty of Severe Accidents” (MUSA) EURATOM project. In particular, results from a training exercise, involving an exploratory UQ application with a simplified SA scenario (namely the Phebus FPT1 test) and a limited number of input parameters, are employed as a basis for the performed sensitivity analysis. As a result of the application of the different sensitivity methods, the parameters, or “features”, selected by each one of them are reported and documented. Additionally, a thorough cross-comparison of the obtained results is conducted in the attempt to work out an optimization of the sensitivity analysis process.