For deep geological repositories for radioactive waste, the release of contaminants to the biosphere has to be assessed in advance, which can only be done by modelling all relevant effects in an integrated, coupled model. Such computation models are typically rather complex, as they combine a lot of influences from various processes. As a result, they often show a highly non-linear behavior.

As there are many uncertainties influencing the calculation results, sensitivity analysis is an important tool for investigating the model behavior. It is not only appropriate for directing research activities, but can contribute essentially to a proper model understanding and even reveal errors in the model or the data.

Uncritical application of standard methods can lead to correct sensitivity estimation, but it is also possible that it yields misleading results, jeopardizing the benefit of sensitivity analysis. The research program presented here was set up in order to investigate standard methods as well as new developments in sensitivity analysis and test their applicability to performance assessment model results. The final goal of the investigations was to provide some guidance to a modeler for performing an effective and meaningful sensitivity analysis.

Three performance assessment models, describing hypothetical repositories for radioactive waste in different geological formations, were defined. These models show different particularities that are typical for their specific type, like output results spread over many orders of magnitude, occurrence of a considerable number of zero-runs, a two-split output distribution or a nearly non-continuous behavior. For each model a set of uncertain input parameters with plausible probability density functions (pdfs) was defined.

The models were calculated numerous times, using parameter samples of different sizes and based on different sampling algorithms. A variety of different standard and advanced methods of numerical sensitivity analysis were applied. Some experiments were done with correlated input parameters and transformation of model output. Moreover, graphical methods of sensitivity analysis were applied.

The assessment was oriented at the following questions: - How robust are the results? Do they considerably change if a different sample of same size is used? How many runs are necessary to achieve stable results? - Do the different sensitivity measures and graphical methods qualitatively agree about the main sensitivities? - Are the results plausible and understandable? - Are all sensitivities detected by the different methods? - Which sampling algorithm seems best? - Can the significance of sensitivity analysis be improved by transforming the model output to a more appropriate scale? - How numerically effective are the different methods of sensitivity analysis?