Application of machine learning for identifying plutonium diversion in molten salt reactors

Abstract

Molten salt reactors (MSRs) have advantages over traditional light water reactors; however, nuclear material safeguards remain an obstacle to their deployment. The nontraditional liquid fuel of the MSR designs examined in this work and additional material flow streams create challenges for safeguards. These properties not only preclude traditional item-counting safeguards but also create new material diversion pathways. The approach proposed here involves modeling a reactor similar to currently proposed molten salt reactors using Serpent and SCALE, simulating its operation under normal conditions and under various scenarios of plutonium removal. SCALE Sampler is used to introduce isotope-specific nuclear data uncertainty perturbations across 10,000 reactor operations for each scenario. Gamma-ray, neutron and alpha spectra measurement signatures are then collected from each perturbed reactor simulation to identify differences caused by plutonium removal. These signatures are then passed to a machine learning model that classifies each dataset as originating from either normal reactor operation or reactor operation during which plutonium diversion occurred. The highest performing developed machine learning model is capable of correctly classifying over 90% of all reactor simulations for most examined diversion scenarios. Alpha spectra features created by 242 Cm, 239 Pu, 240 Pu, along with gamma-ray spectra features created by 239 Np, 133 Xe, 132 I and other isotopes, proved instrumental for the classification process. The results indicate that, given appropriate measurement signatures, plutonium diversion can be recognized with confidence, complementing the proposed isolate-and-contain safeguards approach for MSRs. Accordingly, MSRs would particularly benefit from safeguards by design that enable collection of these signatures.