Modeling of critical experiments and its impact on integral covariance matrices and correlation coefficients

TL;DR

This paper investigates how different modeling assumptions of experimental data significantly influence the resulting covariance and correlation matrices, impacting nuclear data validation and uncertainty analysis.

Contribution

It demonstrates the substantial effect of modeling choices on covariance matrices and correlation coefficients in nuclear experiments, highlighting the importance of accurate modeling in validation procedures.

Findings

Correlation coefficients vary from 0 to 1 depending on modeling assumptions.

Different modeling assumptions lead to significantly different covariance matrices.

Choice of modeling impacts validation bias and uncertainty quantification.

Abstract

In this manuscript we study the modeling of experimental data and its impact on the resulting integral experimental covariance and correlation matrices. By investigating a set of three low enriched and water moderated UO2 fuel rod arrays we found that modeling the same set of data with different, yet reasonable assumptions concerning the fuel rod composition and its geometric properties leads to significantly different covariance matrices or correlation coefficients. Following a Monte Carlo sampling approach, we show for nine different modeling assumptions the corresponding correlation coefficients and sensitivity profiles for each pair of the effective neutron multiplication factor keff. Within the 95% confidence interval the correlation coefficients vary from 0 to 1, depending on the modeling assumptions. Our findings show that the choice of modeling can have a huge impact on integral experimental covariance matrices. When the latter are used in a validation procedure to derive a bias, this procedure can be affected by the choice of modeling assumptions, too. The correct consideration of correlated data seems to be inevitable if the experimental data in a validation procedure is limited or one cannot rely on a sufficient number of uncorrelated data sets, e.g. from different laboratories using different setups etc.