Statistical learning and inverse uncertainty quantification in nuclear thermal-hydraulic simulation: application to the condensation modelling at the safety injection

The BEPU (Best Estimate Plus Uncertainty) methodology is based on the validation and the uncertainty quantification of the physical models used in the nuclear computer codes. Having a robust methodology to calibrate a physical model, validate it and quantify its uncertainty is a major challenge. Therefore, the objective of this thesis is to develop methodologies for the validation and uncertainties quantification and apply them to the practical case of the condensation model at the Emergency Core Cooling (ECC) system. The first step was the development of a new Bayesian framework for the calibration and inverse uncertainty quantification of physical models using a multiplicative model uncertainty. The second step was a detailed thermohydraulic analysis of the experimental data in order to improve the physical understanding of the ECC condensation in the cold leg of a pressurised water reactor during a Loss Of Coolant Accident (LOCA). Two types of condensations occur in the cold leg: the condensation at the ECC and the condensation downstream the jet. The temperature at the outlet of the cold leg, and therefore at the entrance of the reactor's core, depends on these two models. The Bayesian framework is applied to calibrate a new condensation model at the ECC. This model has been validated and the uncertainties have been quantified. In the last step, an extension of the CIRCE methodology (called CIRCE 2-Steps) was developed. CIRCE is a methodology to determine the model uncertainty based on the difference between the experimental and calculated value of a quantity of interest. The proposed extension improves the estimation of the uncertainties of several models combined together and interacting on the same output variable, such as the temperature at the end of the cold leg. (author)