A comprehensive Bayesian framework for the development, validation and uncertainty quantification of thermal-hydraulic models

Abstract

The development, validation and uncertainty quantification of closure laws used into thermal–hydraulic system codes is a key issue before applying the BEPU (Best Estimate Plus Uncertainty) methodology for safety studies and licensing of nuclear reactors. The assessment of those physical models requires tuning some parameters against available experimental data. This paper presents a methodology called Bayesian calibration, which allows a more robust and reliable assessment, selection and uncertainty quantification of physical models.In this work, the experimental and predicted values are linked by means of a multiplicative random variable which represents the model uncertainty. This hypothesis is suitable for models scaling across many orders of magnitude as it is the case in nuclear thermal–hydraulic. Several empirical models, which describe the same physics with different formalisations, are calibrated. Then, the one which is the best-suited according to different statistical indicators is selected. A statistical validation is performed with a Leave One Out (LOO) technique which allows using the same database for both assessment and validation of the physical model. Finally, the uncertainty on the selected best-fitting model is assessed.This framework is applied to condensation heat transfer correlations for safety injection based on the COSI (COndensation at Safety Injection) experiments.