Development of Machine Learning Algorithms for Directional Gamma Ray Detection

Directional detection is an important component in the search for rouge radioactive materials, but success is often burdened by Poisson statistics, scatter, and limited count times. Thus, it is necessary to utilize algorithms that can maximize the confidence of conclusions drawn from data associated with large variability. A typical method of directional detection is to create a prepopulated database of detector responses with known source locations. An unknown detector response is then compared to this database by preforming a least squares assessment to estimate the angle. This method is limited when searching for sources at distances and environments considerably different than those available in the database. A method capable of analyzing data with large amounts of variability will advance directional detection capabilities. To that end, a residual neural network was implemented on a series of simulated unknown source location scenarios to develop an algorithm for search applications less dependent on source distance. Monte Carlo Neutral Particle (MCNP6) was used to simulate detector responses for an array of 8 5x10x41 cm NaI detectors in a cylindrical configuration. A large dataset was simulated with distances ranging from 1-15 meters at random angles on a plane. Simulated source settings were set to represent a 10 second count of a 1.4 micro-Curie colbalt-60 point source. These conditions yielded various uncertainties, some over 10%, emulating the limited statistics of real-world scenarios and greatly complicating the task at hand. Current results for the first implementation of this neural network have yielded a correct location within 3 degrees for approximately 50% of trials with an overall average angular error of 4.5 degrees, compared to the least squared method which yielded a correct location within 3 degrees for approximately 44% of trials with an overall average angular error of 4.9 degrees. This work serves to investigate machine learning architectures for the use of directional detection problems, in order to create an algorithm effective over a reasonably large search area. The work will present more accurate simulations of the scatter and obstructions of real word environments, benchmarking against experimental measurements, and investigating various algorithm and detector array designs.