Machine Learning Models for Molten Salt Reactor Safeguards
Open Access
- Author:
- Kovacevic, Branko
- Graduate Program:
- Nuclear Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- January 29, 2025
- Committee Members:
- Jon Schwantes, Program Head/Chair
Amanda Johnsen, Chair & Dissertation Advisor
Marek Flaska, Major Field Member
Aleksandra Slavkovic, Outside Unit & Field Member
Xing Wang, Major Field Member
William Walters, Special Member - Keywords:
- Molten salt reactors
Machine learning
safeguards
gamma-ray modeling
neutron modeling
alpha spectra modeling - Abstract:
- As some molten salt reactor (MSR) concepts approach certification and eventual deployment, their safeguards methodology needs to be defined, and methods of safeguards implementation created. The liquid state of the fuel in some types of MSRs prevents the use of traditional material accountancy methods, such as counting individual fuel elements or assemblies; some proposed MSR safeguards approaches rely only on containment and surveillance. The research presented here hypothesizes that it is possible to detect material diversion from liquid-fueled MSRs by monitoring measurement signatures of samples taken from a reactor. The approach is to model a representative MSR, implement multiple plutonium diversion scenarios within the model, and then leverage machine learning algorithms to synthesize a suite of potential diversion signatures to examine the likelihood of diversion detection. As the idea is to focus on less resource-intensive measurements to simplify practical safeguards, the modeled signatures come from gamma-ray and alpha spectrometry and neutron counting. Although this approach is implemented on a single MSR design, it aims to create a process methodology that can be exported to other liquid-fueled MSR designs. To explore this methodology, a molten salt reactor model fueled with low enriched uranium was developed using Serpent, a Monte Carlo-based transport code, and SCALE, a deterministic transport code. The SCALE module Sampler was used to introduce isotope-specific nuclear data uncertainty perturbations across 10,000 reactor operations for each operational scenario. The models were used to obtain resulting isotope concentrations for both a reference case and for reactor operations that experienced one significant quantity (SQ) of plutonium diversion over varying protracted and abrupt time frames. The machine learning algorithms were also used to identify isotope combinations whose concentrations, when used to create a dataset, indicate the 1 SQ plutonium removal with high confidence. Additionally, noise was introduced to the isotope concentrations to simulate measurement uncertainty, enabling the determination of target assay uncertainty values for the identified isotopes. As the required assay uncertainties were very low, the resulting isotopic concentrations were then used to model emission signatures that might improve diversion detection capabilities. Gamma-ray spectra modeled in GADRAS at various stages of sample decay reveal a multitude of characteristic peaks indicating material diversion. Neutron signatures modeled in the SCALE module Origen and in the Monte Carlo N-Particle (MCNP) code exhibit differences in total neutron counts comparable to plutonium concentration differences between diversion and reference cases. Alpha spectra signatures, modeled in MCNP, provide the only modeled measurement features created by plutonium isotopes, greatly aiding the machine learning algorithms in their classifications of reactor operation cases. The machine learning algorithms achieved better than 88% end-of-cycle classification accuracy between the reference and all but one examined 1 SQ plutonium diversion cases when using an optimized selection of features from the measurement signatures. The end-of-cycle classification of early-cycle abrupt diversions represents a greater challenge, but such diversions are likely to be detected either through earlier measurement signatures or through reactor operational parameters at the time of diversion.