Improved methodology for the determination of model uncertainties using the example of ATHLET

Abstract

In the framework of the project RS1597 several aspects of carrying out an inverse uncertainty analysis have been addressed. As the question of parameter choice of interest (parameters related to the reflooding phenomena) has been set at the beginning of the project, the first step of the SAPIUM guidelines have been already fulfilled. As this phenomenon served as the basis for several international projects previously, the set of experiments have been already tested and have been deemed adequate and appropriate for the quantification process.

The approach chosen and implemented for our purposes is the so-called ABC (Approx-imate Bayesian Computation) method. This method circumvents the derivation of the likelihood function by introducing a rejection scheme based on an appropriately chosen distance metric comparing the simulated values with the observational data, or their properties (mean, skewness, etc.). Several implementation approaches have been looked into, and in the end, Python has been selected due to its flexibility, adaptability, extensive list of available libraries, and for its available easy connection to other GRS tools (MCDET, SUSA) addressing aspects of (forward) uncertainty quantification. The method had been implemented and tested on small examples before moving to adapt it to ATHLET simulations. The last steps of the approach (verification and validation) involve a forward uncertainty analysis applying the derived distributions of the uncertain input parameters.

For the validation of AC²/ATHLET using SET and CET, in the frame of the current project reflooding tests of the test series FEBA, FLECHT and PERICLES were investigated.

Additionally, the ATHLET validation manual has been extended with additional guidance on the application of Wilks’ formula, the determination of sample sizes for uncertainty analysis and the treatment of code crashes, first for ATHLET 3.3 and then extended again for ATHLET 3.4. In the frame of international activities GRS participates mainly in two projects resp. networks, OECD/NEA ATRIUM and FONESYS. Based on the current status of modelling for the inverse uncertainty quantification used for AC²/ATHLET applications, some topics for improvement could be identified: investigating different ABC metrics, expanding the experimental data base, saving chain-information mid analysis, including modelling bias, better resource allocation, the question of the likelihood function as well as surrogate modelling.