The safety of nuclear waste repositories can be assessed by analyzing all relevant Features, Events and Processes (FEPs) (Savage 1995) and their interactions in order to generate scenarios that support the assessment of radiological consequences (Campbell & Cranwell 1988). Yet, ensuring the comprehensiveness (STUK 2015) of this process of Scenario Analysis can be difficult due to time and budget constraints. For instance, it may not be possible to analyze all scenarios or, more modestly, to identify which scenarios need to be analyzed to conclusively assess the safety of the repository. Moreover, analysts have to characterize epistemic uncertainties, too (Aven & Zio 2011).
To address these challenges, we propose a method of Scenario Analysis based on Bayesian networks (BNs) (Pearl & Russell 2003) whose nodes correspond to the FEPs, which are modeled as stochastic variables with finite discrete states. Interactions between the FEPs, shown as links between these nodes, are modeled through conditional probabilities that are derived either by (i) reproducing the underlying phenomena (such as copper corrosion and radionuclide transport) through computer-code simulations or laboratory experiments, or (ii) eliciting expert judgments (Laitila & Virtanen 2016). Each scenario, defined as a complete combination of FEP states, leads to unacceptable radiological consequences with its corresponding failure probability.
Because comprehensiveness can be difficult to guarantee due to computational or experimental resources, we propose an iterative algorithm for guiding the selection of the scenarios for which simulations need to be run for assessing whether or not the failure probability exceeds the stated acceptability threshold. At each iteration, the algorithm (i) guides the choice of the next simulation and (ii) uses the result to update the interval-valued failure probability. The algorithm terminates when the interval becomes conclusive, either because its lower bound exceeds the acceptability threshold (repository unsafe) or because its upper bound is below the threshold (repository safe).
In the BN, epistemic uncertainty is dealt with by aggregating probability estimates from different experts through weighting, and by using multilinear optimization (Toppila & Salo 2013) to determine conservative estimates of the bounds of the failure-probability interval.
We illustrate our method by building a BN for a simplified nuclear waste repository in which most interactions are addressed through expert judgments while a physical model of radionuclide transport (Cadini et al. 2016) is employed to simulate the release rate. In this example, the safety assessment can be completed by running simulations for no more than 25 % of all scenarios while still guaranteeing conservativeness in the treatment of epistemic uncertainties.