Year
2024
Abstract
| The liquid nature of molten salt reactor (MSR) fuel and its anticipated online processing schemes open new pathways for material diversion while precluding traditional material accountancy measures such as counting discrete fuel elements. To enable the identification of signatures indicating plutonium diversion from uranium-fueled MSRs, we created a reactor model in Serpent based on the Molten Salt Demonstration Reactor but adjusted the enrichment, fuel cycle (8 years), and fuel feed corresponding to reactor types currently under development.The modeled reactor’s operation was simulated without any material removal for reference, and with various plutonium diversion scenarios. Subsequently, we modeled various measurement signatures with priority on non-destructive methods, and evaluated differences between diversion and reference cases, aided by machine learning (ML) models. The Gamma Detector Response and Analysis Software (GADRAS) was used to model gamma-ray signatures using samples from various material streams and sample decay steps. MCNP and the SCALE package ORIGEN were used to model the neutron emissions from fuel salt samples removed from the reactor at various operational steps. MCNP was used to model fuel salt samples’ alpha detector response. Signatures were modelled by simulating the most influential detection physics such as detector dead time, Poisson counting noise, and Gaussian peak broadening. These signatures were created for each perturbed reactor operation, which were then collected into reference and diversion datasets whose classification by a machine learning algorithm was used to signal material diversion.ML models were particularly useful for synthesizing different signatures to achieve detection of plutonium diversion while accounting for various forms of uncertainty. The Sampler SCALE package was used to introduce nuclear data uncertainty (cross section, decay data and fission yield) to each isotope within the reactor model. Reactor operation was then modelled 10,000 times with different perturbation factors, yielding a distribution of concentrations, specific to each isotope based on its nuclear data uncertainty, creating multiple reactor operation outcomes to be used for machine learning datasets.One examined diversion scenario is a constant rate, protracted removal of 1 SQ of plutonium over the entire reactor cycle. The scenario considers a perfectly executed diversion where fuel salt is removed from the reactor, reprocessed to remove only plutonium, and return all other constituents to the reactor. While performing such a diversion would be very challenging, it was determined to be the most challenging to detect, as a less perfect diversion would remove additional fission products and actinides. The ML model, when provided with a mix of a fuel sample’s selected gamma-ray signatures, total neutron count, and alpha spectra, correctly classified 1 SQ protracted plutonium diversion in 87.7 +/- 0.9% of all perturbed reactor operations.Additionally, the isotopic compositions of various material streams within the reactor, with added noise to simulate measurement uncertainty, were used in ML models to classify reference and diversion cases. The goal was to identify isotopes whose concentrations are more likely to indicate plutonium diversion with high accuracy, but also to evaluate how low measurement uncertainties must be to deliver reasonable classification accuracy, informing future experimental efforts. |