A machine learning approach to predict radioxenon isotopes concentrations using experimental data

Abstract

Monitoring atmospheric radioxenon isotope concentration is crucial for verifying compliance with the Comprehensive Nuclear-Test-Ban Treaty (CTBT). The most sensitive method for radioxenon isotope detection is beta-gamma coincidence measurement. Quantities of activity concentrations and, more specifically, their ratios are vital for analyzing spectra from atmospheric radioactivity to characterize the emission source. This study aims to develop a novel approach utilizing a machine learning algorithm to estimate the activity concentrations of radioxenon isotopes from real spectra in the format of experimental two-dimensional raw spectra without the need for extensive mathematical calculations for subtracting interferences and backgrounds. This method can be applied as a pre-screening method in a fast way to highlight spectra with concentration anomalies and help analysts in the interactive review step. Traditional methods like region-of-interest (ROI) and simultaneous decomposition analysis tool (SDAT) have been used to estimate net counts for each specific ROIs. However, this study employs deep convolutional neural networks (CNNs) to extract fine structure and patterns from the complex spectra, capturing both local and global features from entire beta-gamma coincidence spectra without relying on predefined ROIs.