Review Of Methods To Aggregate Diversion Detection Probabilities Across Multiple Material Strata

Statistical models are used to calculate probability of detection of diverted material for individual nuclear material strata to assess the effectiveness of IAEA verification inspections. The safeguards verifications use stratified inventories or materials flows, whereby the material is grouped into strata on the basis of similar physical and chemical characteristics. The detection probabilities (DPs) can be aggregated across material strata to determine the probability of detecting material diversion at the facility level. To aggregate DP to the facility level, one must account for all realistic ways in which a total amount of diverted material may be split among different strata. A brute-force enumeration to calculate aggregate detection probability (ADP), referred to as Partitions method, considers all possible combinations of splitting the goal amount among the existing strata. The algorithm then finds the combination that meets a specified criterion, such as minimum ADP in a facility across all strata. The Partitions method finds the global minimum of the ADP because it explicitly considers all different combinations. However, when considering a large facility with several strata the algorithm comes with a high computational cost. The computational time is observed to increase exponentially with increasing number of stratums. Another approach to compute ADP uses Pareto frontiers, inductively updating the frontier by evaluating combinations with one additional stratum at each inductive step. The process iterates until the final stratum is included to calculate the final ADP result. We have studied two other methods, which we refer to as a modified Greedy Algorithm and Guided Genetic Algorithm. The paper will discuss simulation results for these aggregation algorithms in application to representative mock inspection data and facilities. The simulation results presented in this paper complement the work described concurrently in the INMM/ESARDA conference paper by Bevill et al. (2021).