The Development of Advanced Analysis Techniques for Gamma-Ray Spectra
Restricted (Penn State Only)
- Author:
- Fjeldsted, Aaron
- Graduate Program:
- Nuclear Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- July 16, 2024
- Committee Members:
- Dipanjan Pan, Professor in Charge/Director of Graduate Studies
Marek Flaska, Major Field Member
Kenan Unlu, Major Field Member
Azaree Lintereur, Special Member
Douglas Wolfe, Chair & Dissertation Advisor
Wesley Reinhart, Outside Unit & Field Member
Patrick Lofy, Special Member - Keywords:
- Nuclear Security
Gamma-Ray Spectroscopy
Machine Learning - Abstract:
- Nefarious actors can unleash devastating societal consequences if they successfully obtain radiological or nuclear material. Even the accidental mishandling and loss of radiological sources are of concern as this can lead to significant health risks and a lack of trust in nuclear-related industries. The majority of the radiological and nuclear materials of concern in these trafficking or hazardous scenarios emit radiation in the form of gamma-rays. Detectors sensitive to gamma-rays are leveraged to collect spectral data which ascertain radiological constituents in the environment. However, the accurate assessment of a radiological environment can be complicated due to many factors, such as the presence of multiple gamma-emitting radioisotopes, background radiation, shielding occlusion between a source and detector, and limited measurement times. These factors can convolute the spectral shape, which is the key to extracting accurate radiological situational awareness. These challenges have led researchers to pursue machine learning techniques in the analysis of gamma-ray spectral data for their data-centric pattern recognition and classification capabilities. Even with the powerful analytical tools afforded by machine learning-based spectral analysis, novel challenges continue to arise. For instance, many studies have implemented neural networks for their analysis; however, the process by which neural networks identify radioisotopes can at times be hard to interpret, a quality that is worrisome for high-stakes applications. Furthermore, relatively little effort has been dedicated to developing robust algorithmic approaches for scenarios where testing data diverges from training data. Additionally, there has been insufficient focus on the algorithmic and dataset enhancements required to create reliable machine learning models for gamma-ray spectral analysis. This dissertation advances the current literature on automated gamma-ray spectral analysis through the development of innovative algorithmic and data synthesis techniques. These novel methodologies place significant importance on the gamma-ray spectra, relying on them for extracting spectral signatures and assessing radiological environments. In terms of algorithm developments, various feature extraction methods were investigated to identify the most important spectral regions for isotope identification. This knowledge was then leveraged in the formulation of a novel feature-driven analytical approach whereby the model-derived feature importance values were used in the detection of out-of-distribution sources. Additionally, a principled statistical measure was developed to quantify the similarity between measurement configurations. This measure was useful in contextualizing changes in isotope identification capabilities due to differences between training and testing data, as well as the development of a novel dataset generation methodology capable of identifying spectrally optimized measurement parameters for dataset procurement. The research presented in this dissertation holds significant value for the field of automated gamma spectroscopy as it introduces and validates spectrally centric methodologies aiming to enhance the robustness and accuracy of spectral analysis.