In this study, using a machine learning method, in particular, a deep learning approach, we propose for the first time a model of Beta-Gamma coincidence radioxenon spectra classification. Specifically, by means of real data from the noble gas system in Charlottesville (USX75) as part of the International Monitoring System (IMS) operated by CTBTO between 2012 and 2019, we apply the convolution neural network (CNN) technique based on the absolute concentration of each radioxenon isotope. The results show that without utilizing background spectra, interference corrections, and without determining the activity concentration of each isotope, the automatic classification can be carried out with high accuracy. This implies that categorization through deep learning does not require the knowledge of screening threshold values that are applied for sample categorization after applying the Net Count Calculation (NCC) analysis method used currently by the International Data Centre (IDC) of the CTBTO. Our results support that by synthesizing nuclear engineering and deep learning disciplines, experts can accelerate and optimize the review process of background and CTBT-relevant samples by an average accuracy of 92% and 98% respectively.
Classification of Beta-Gamma coincidence raw radioxenon spectra by deep learning (CNN technique) as pre-screening for CTBT relevant samples.