Deep Learning driven analysis of gamma ray spectra of nuclear material

Nuclear safeguards inspectors use gamma-ray spectroscopy extensively, for example for determining uranium or plutonium isotopic composition in items or for verifying spent nuclear fuel. Current analysis methods detect and measure energy peaks specific to the expected radioisotopes and exploit the ratio of peak areas (based on defined region of interest) to determine the isotopic composition. These methods provide accurate results for many applications, but they also have limitations in cases of complex spectra and they are supervised methods. Over the last years, deep learning has revolutionized the analysis of images and other sensor data. It can be applied to various tasks, including detection, recognition, classification, data enhancement (super-resolution, de-noising) and data generation. This paper explores the potential of deep learning for analyzing gamma spectra of nuclear material (e.g., enriched uranium) for isotopic composition determination. Both experimental data and simulated gamma spectra are used to provide the training data required for deep-learning-based gamma spectrum analysis. In the paper, we report on the evaluation of deep algorithms for analysis of gamma-ray spectra and their applicability in nuclear safeguards and non-proliferation applications.