Neural networks for the analysis of 2D radio-xenon beta gamma spectra

Abstract

Atmospheric traces of radioactive xenon can be a strong indicator for underground nuclear fission reactions. 131mXe, 133Xe, 133mXe and 135Xe are the primary gaseous isotopes/isomers currently used to identify and classify nuclear events. During decay, each of these radioactive species produces a unique beta-gamma energy spectra, which can be measured using beta-gamma coincidence counting. Current operational Xe beta-gamma spectrum analysis software relies on Region of Interest (ROI) counting (Bowyer et al. in J Environ Radioact 59(2):139–151, 2002). This algorithm occasionally produces mismeasurements, especially when quantifying meta-stable isomers, due to overlapping ROIs and shifts in detector calibration in fielded systems over time (Ringbom and Axelsson in Appl Radiat Isot 156:108950, 2020). In an attempt to better de-convolve overlapping isotope spectra we have developed a technique that applies a supervised neural-network implemented in TensorFlow with Keras to classify and quantify the isotopes and mixtures of isomers based on their beta-gamma spectra—similar to image recognition. From this, we have improved upon the false-positive rate for classification and regression models, however challenges remain with dealing with differing detector energy calibrations and with estimating measurement uncertainty.