Neural Assessment Of Non-destructive Assay For Material Accountancy

The IAEA spends significant effort verifying the non-diversion of nuclear materials at bulk handling facilities such as reprocessing plants and uranium enrichment plants. This verification often requires that the IAEA collect and analyze dozens of small samples taken from the process to directly confirm the concentration of uranium and/or plutonium in the process materials. The collection of these samples is often resource intensive for both the IAEA and facility operator and the subsequent IAEA analysis is quite expensive. Given the increasing amounts of nuclear material under IAEA safeguards and their fixed budget, significant improvements to IAEA resource could be achieved if these samples could be reduced while still reaching the needed diversion detection probability. While running a nuclear facility the operator collects large amounts of process monitoring data. This data is already used for in a few limited safeguards purposes applications and has been combined with non-destructive assay (e.g. gamma measurements on the material) to provide material flow monitoring. However, the current emphasis favors destructive analysis over process monitoring because the measurement uncertainty is significantly less and therefore it is easier to meet the detection probability metrics required by the specific safeguards approach. This work seeks to improve the power of analysis that can be performed via process monitoring data streams such that they can be used to supplement or replace destructive assay in some instances. The simulated process data used for this project representing the flow of nuclear materials through a hypothetical aqueous reprocessing plant was produced by the Separation and Safeguards Performance Model. The model is based in MATLAB and provides at control points throughout the process. Hundreds of simulations, representing both nominal and diversion scenarios, were used to train a transformer-based model on the system dynamics. This training allowed the transformer model to predict the movements of materials and compare the predicted behavior to the true behavior. Deviations from the predictions of a well-calibrated model were then used to detect diversion in a hold-out set of simulations with results that often exceeded the discrimination power of Pagh’s test exploiting sample results on the same data.