Sensitivity of the WIPP performance assessment calculations to model parameters - 073

The Waste Isolation Pilot Plant (WIPP) is located in southeastern New Mexico and operated by the U.S. Department of Energy (DOE) as a disposal facility for transuranic (TRU) waste. The WIPP must comply with various environmental regulations regarding containment of the waste. Some of these regulations are framed in terms of the probabilities of releases to the accessible environment across all pathways (total releases) over 10,000 years. The accessible environment is defined as the surface and the lateral boundaries of the WIPP site. The potential impacts of potash mining near the WIPP and inadvertent human intrusion into the repository by exploratory drilling must be taken into account in these analyses. Sandia National Laboratories (SNL) conducts performance assessments (PAs) of the WIPP using a system of computer codes in order to demonstrate compliance with this regulation. The WIPP PA model system is a coupled collection of 24 process-oriented models that are used to predict chemical and physical changes in the repository, radionuclide transport within the repository and to the surface via fluid flow, and subsurface migration of radionuclides to the lateral boundaries of the accessible environment. These process-based models are typically run for a number of specific intrusion scenarios, where an intrusion scenario is defined in terms of the time of an intrusion, the general location of an intrusion, whether a brine pocket is penetrated, and the state of the repository at the time of the intrusion. The state of the repository is determined by whether it was previously un-intruded, intruded without encountering a pressurized brine pocket, or intruded with the introduction of pressurized brine. There are no stochastic events modeled within these models. However, many parameters of these models are assumed to have epistemic uncertainty about their values. Distributions were assigned to the parameters using subjective techniques. The uncertainty in the parameters was propagated to the results using Monte Carlo methods. Specifically, three independent replicates of 100 sets of parameters were generated using Latin Hypercube Sampling (LHS). LHS helps to ensure that the distributions were adequately sampled using only 100 trials per replicate. Each of the process-level models was then run using each of the 300 sets of parameters for each of the scenarios relevant to that model. The results of these simulations were stored and subsequently used as input to the model CCDFGF. CCDFGF is used to simulate a series of potential futures for the repository. Each of these futures represents a simulation covering 10,000 years following the closure of the repository. Six stochastic processes are invoked to generate a unique 'history' for each simulation: the location and timing of drilling, the occurrence and impact of mining intrusions, the penetration of contact- or remote-handled waste, the selection of the waste streams encountered by a borehole, the penetration of brine pockets, and the plugging of boreholes. CCDFGF relies on the results computed by the process-oriented models to compute releases from the repository. Ten thousand futures are generated by CCDFGF for each of the 300 sets of model parameters used in the analysis. The 10,000 realizations are used to construct a distribution of potential releases for each set of parameters. Thus this analysis produces 300 probability distributions for each output variable. Global sensitivity analyses were conducted on the results from CCDFGF. The purpose of the global sensitivity analysis was to assess the effect that a parameter has on an output variable over the range of values exhibited in the LHS samples. Because the output for each set of parameter values was a distribution of possible releases, the impact of the parameters on the mean releases was evaluated. Fifty-six parameters were evaluated in the current analysis. Stepwise linear multiple regression using the ranks of the parameters and output variables was used to evaluate the relative importance of the various sampled parameters on the estimates of potential releases. The rank of a value is an integer representing its relative position in the sorted list of the values. Regression on the ranks of the data tends to linearize monotonic response curves and standardizes the variability in the outputs and parameters by mapping the data into identical ranges. Scatter plots of the dependent versus independent ranked variables resulting from the analysis were examined to determine if there were any obvious non-monotonic relationships. Partial correlation coefficients are used to order the parameters with respect to their impact on model responses. Cuttings, cavings, spallings and direct brine releases account for an overwhelming majority of the total releases. Nearly 90% of the uncertainty in total normalized releases is due to uncertainty in waste shear strength. Waste shear strength affects the mechanical integrity of the waste surrounding the borehole. The second most important parameter contributing to variability in total releases accounts for only about 2% of the response. These results and the sensitivities of the individual release pathways will be discussed and compared to previous analyses