Citation:
P. O’Neal, “Development and Verification of a Nuclear Forensics Methodology for the Attribution of Plutonium Using Data Science Methods”, Ph.D. Dissertation, Nuclear Engineering, Texas A&M University, College Station, TX (2024).
Abstract:
An advantage the global community has in preventing nuclear terrorism is the difficulty for a nonstate actor to procure special nuclear material (SNM). The regulation of SNM is essential in stymying adversaries. A nuclear forensics methodology, able to determine the provenance of SNM, like plutonium (Pu), will aid the international community in deterring nuclear smuggling. If Pu is recovered outside of regulatory control, an attribution capability would help inform conventional investigators. In both cases of theft and state hand-off of Pu, a guarantee that an offending party would be discovered and punished could force preemptive abandonment of any planned misdeeds. The goal of this research was to develop and verify a nuclear forensics methodology for attributing unknown separated Pu samples using machine learning techniques. The methodology needed to be capable of identifying the following three attributes: the reactor-type that produced the Pu sample, the burnup of the irradiated uranium fuel that produced the Pu sample, and the time since irradiation (TSI). The methodology also needed to be robust enough to attribute samples that contain a mixture of Pu from multiple different reactor sources. A set of isotope ratios was used as the forensics signature and the training of the machine learning models utilized data from a library of Monte Carlo reactor neutronic and fuel burnup simulations. Lastly, the methodology needed to be validated by demonstrating that it could successfully attribute physical Pu samples.rnResearch proceeded in three main parts. First, machine learning models suitable for this application were identified, and were then trained and tested for attributing single reactor type Pu samples to assess feasibility. Second, the methodology was validated with a Pu sample separated from low enriched uranium dioxide (LEUO2) irradiated in a thermal neutron flux spectrum. Third, the machine learning methodology was adapted to attribute samples that were sourced from multiple reactor types. Additionally, a method for estimating the machine learning models’ prediction uncertainty that considered the Pu sample’s measurement uncertainty was investigated. Ultimately, all main objectives were successfully achieved. This is the first example in open literature of a methodology for attributing mixed reactor type Pu samples.