Application of Machine Learning Techniques for Radioactive Material Localization with an Array of NaI Detectors
Open Access
- Author:
- Durbin, Matthew
- Graduate Program:
- Nuclear Engineering (PHD)
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 12, 2022
- Committee Members:
- William Walters, Major Field Member
Patrick Mcdaniel, Outside Unit & Field Member
Marek Flaska, Major Field Member
Azaree Lintereur, Chair & Dissertation Advisor
Arthur Motta, Professor in Charge/Director of Graduate Studies - Keywords:
- Radiation Detection
Radiation Localization
Nuclear Security - Abstract:
- Radioactive materials have numerous applications in energy, medicine, and industry, and are distributed across the globe. For nuclear and homeland security purposes, it is necessary to have technologies capable of localizing potentially rogue radioactive sources. One modality of interest is directional detection. By analyzing the distribution of counts across an array of stationary gamma ray detectors, it is possible to predict the angle (direction) of a gamma ray source relative to that array. This research investigated leveraging machine learning (ML) techniques to improve the utility of directional detection using an array of NaI detectors. Results with simulated data showed ML algorithms can outperform other approaches of the literature for a variety of signal-to-noise ratio scenarios and source to array distances. Two key advantages of ML approaches were identified. First, by using domain informed feature engineering, subtle angular dependencies latent within spectral features typically not used in other approaches can be exploited, allowing for additional, physically significant degrees of freedom for ML algorithms to form models. Second, by training on large datasets of representative scenarios, the complex relationship between source location and the array response can be generalized. To investigate the feasibility of use in real applications, a variety of physical dependencies and environmental complexities were considered, including source energy, varying background, and obstructions. For each of these considerations, their impact on the array response was investigated, and both ML and traditional radiation detection techniques were leveraged to enhance localization performance. Tests were also conducted that showed synthetic models can provide meaningful localization predictions on laboratory data, in agreement with similar simulated tests. The methods used for the static scenario were also extended to an analysis capable of localization with a dynamic array. The novel approach combined the feature engineering and training advantages of ML, with the dynamic path generalization of likelihood-based approaches, leading to successful localization for multiple demonstrations.