The impact of spatial variability of hydrogeological parameters - Monte Carlo calculations using SITE-94 data

In this report, several issues related to the probabilistic methodology for performance assessments of repositories for high-level nuclear waste and spent fuel are addressed. Random Monte Carlo sampling is used to make uncertainty analyses for the migration of four nuclides and a decay chain in the geosphere. The nuclides studied are cesium, chlorine, iodine and carbon, and radium from a decay chain. A procedure is developed to take advantage of the information contained in the hydrogeological data obtained from a three-dimensional discrete fracture model as the input data for one-dimensional transport models for use in Monte Carlo calculations. This procedure retains the original correlations between parameters representing different physical entities, namely, between the groundwater flow rate and the hydrodynamic dispersion in fractured rock, in contrast with the approach commonly used that assumes that all parameters supplied for the Monte Carlo calculations are independent of each other. A small program is developed to allow the above-mentioned procedure to be used if the available three-dimensional data are scarce for Monte Carlo calculations. The program allows random sampling of data from the 3-D data distribution in the hydrogeological calculations. The impact of correlations between the groundwater flow and the hydrodynamic dispersion on the uncertainty associated with the output distribution of the radionuclides' peak releases is studied. It is shown that for the SITE-94 data, this impact can be disregarded. A global sensitivity analysis is also performed on the peak releases of the radionuclides studied. The results of these sensitivity analyses, using several known statistical methods, show discrepancies that are attributed to the limitations of these methods. The reason for the difficulties is to be found in the complexity of the models needed for the predictions of radionuclide migration, models that deliver results covering variation of several orders of magnitude. Correlations between parameters also make it difficult to separate the contribution from each parameter on the output. Finally, it is concluded that even in cases where correlations between parameters can be disregarded for the sake of the uncertainty analysis, they cannot be disregarded in the sensitivity analysis of the results. A new approach for global sensitivity analysis based on neural networks has been developed and tested on results for the peak releases of caesium. Promising results have been obtained by this method, which is robust and can tackle results from non-linear models even when there are correlations between parameters. This represents a considerable improvement over the capabilities of the commonly used traditional statistical methods