Future arms control treaties could conceivably limit the total number of nuclear warheads and involve onsite inspection of individual nuclear weapons. Verification under such an agreement presents a unique challenge necessitating the need to confirm the presence or absence of special nuclear material without direct access to the material or revealing sensitive information. Passive γ-spectroscopy can be of value in these scenarios, typically along with other measurements, to provide information regarding isotopic composition. Prototype gamma detectors with information barriers have been developed in previous efforts, but they have demonstrated limitations in specific use cases. Since the development of these prototype verification systems, there has been an increase in the utilization of machine learning for the comprehensive analysis of γ-ray spectra, often employed to identify specific isotopes and assess their respective quantities. Many of these studies have focused on machine learning techniques, relying on the Gamma Detector Response and Analysis Software (GADRAS) for the generation of synthetic training data in particular use cases. This work explores the approach of generating an array of synthetic γ-spectra of value to the arms control community for training machine learning algorithms. A High Purity Germanium (HPGe) detector response was simulated for seven unshielded uranium or plutonium spheres using GADRAS, GADRAS-DRF, and MCNP to compare observed count rates. MCNP generated spectra correctly identify proper peak energies, however the existing approach results in count rates that differ from those in GADRAS and GADRAS-DRF. This difference is likely due to the necessity to spawn particle histories in the outermost radius to generate counts using Monte-Carlo, which does not reflect the reality of self-shielding of lower energy photons. While GADRAS-DRF lacks full radiation transport capabilities, peak count rates agree within a few percent of the full version of GADRAS.
Year
2024
Abstract