IMPACT OF CORRELATED DATA IN VALIDATION PROCEDURES

摘要

The increasing demand on accuracy of criticality calculations leads to the question, how to treat correlated data of benchmark experiments correctly. Correlations in the process of validation arise for example if a configuration in a series of experiments shares certain components, e.g. fuel rods. Traditional methods of code validation do not treat correlations explicitly, but give a more conservative estimation of the bias. Including correlations can lead to a more precise estimation. In the work at hand, a method is discussed based on Bayes theorem to include correlations in the experimental data when estimating an application case k_(ef)ff. After a short introduction to the applied method and the demonstration of the general capability of it to predict the K_(eff) correctly, it is shown how the calculated Keff of an application case depends on correlations in experimental data for a simple Toy Model. The impact of ignoring correlations and taking to account lower and higher correlations is shown and discussed. Further results are shown using a database of correlated experimental data from the International Handbook of Criticality Safety Experiments to estimate the bias of two application cases. The experiments were chosen following the suggestion of the OECD/NEA EGUACSA Benchmark phase IV and consists of 21 experiments from the LEU-COMP-THERM series 007 and 039. With the extracted data results the calculated keff values of the application cases are shown and discussed for different examples of correlations. The calculations were done by applying Monte-Carlo sampling methods with the GRS tool SUnCISTT, using ORNL's SCALE 6.1.2. This work discusses a possible way of treating correlations in the recalculation of critical experiments to get a more precise bias estimation of the used code system. It will be shown, that the presented method is capable to treat similarities in experimental setups of data used to calculate the bias correctly, independent of the size of similarities. The presented method is in principle capable of assessing regimes with little experimental data to estimate a bias.