A Multi-Step Machine Learning Approach to Directional Gamma Ray Detection

Directional detection predicts the angular component of a gamma ray source's location by analyzing the distribution of counts received across an array of stationary detectors. The array's response to the source is a function of angle, as well as other factors such as distance, energy, and obstructions. The effectiveness of an angular prediction in a real-world environment is therefore dependent on the inclusion of these phenomena when processing the detector array data. With sufficiently representative training data that captures these variables, it is hypothesized that machine learning algorithms can aid in this angular prediction process due to their success in other complex data processing applications. A multi-step approach is introduced, in which machine learning algorithms are tasked with addressing specific complexities of the overall analysis. Initial results indicate that this multi-step method is a viable option which can be used to analyze different components of the array response. Presented here is a proof-of-concept with simulated datasets of three different isotopes, and measured datasets of two different isotopes. Preliminary results indicate that a multi-step machine learning approach improves the overall angle-prediction accuracy compared to a single phase machine learning algorithm and a least-squares comparison to a reference table.