Machine learning for the analysis of 2D radioxenon beta gamma spectra

Clandestine nuclear testing can be detected at a standoff distance using radioxenon beta-gamma analysis. International treaty monitoring organizations depend, in part, upon the activity ratios of various radioxenon types to determine if collected samples are the result of a weapons test or a peaceful purpose such as energy or medical isotope production. However, the currently deployed radioxenon analysis method makes assumptions about the location of energy coincidence counts on a beta-gamma spectrum, such that this method is particularly sensitive to measurement or calibration errors. We propose a machine learning method instead. By exposing a computer algorithm to many representative examples, the resultant computer model detects patterns in the data without making additional assumptions. Both a classification model predicting which radioisotopes are present and a regression model predicting concentrations of the radioisotopes are tested. This work is a proof-of-concept that machine learning can be effectively applied to radioxenon beta-gamma analysis.