Applying deep learning in gamma spectroscopy for radionuclide identification

Abstract

This study presents the results of applying a Deep Convolutional Neural Network model to gamma spectrum classification for radioactive waste management. The approach uses a modified version of the VGG-19 architecture, originally developed for image recognition with 1,000 mutually exclusive classes. In the modified architecture, gamma spectra are used as input, and the model performs nonexclusive classification into ten classes representing the radionuclides most commonly encountered at IPEN's Radioactive Waste Management Department: Am-241, Ba-133, Cd-109, Co-57, Co-60, Cs-137, Eu-152, Mn-54, Na-22, and Pb-210. Gamma spectra were generated using Monte Carlo simulations performed with PENELOPE/PenEasy, simulating an HPGe detector and sources placed inside a steel drum filled with paper, representing the typical contents of drums managed by the department. The dataset was augmented by combining these simulated spectra to generate new spectra containing up to four radionuclides. Several detector-to-drum distances (41 cm, 46 cm, 51 cm, and 56 cm) were used to create a representative dataset. The spectra acquired at 56 cm (150 original spectra, expanded to 375 after the augmentation process) were used for validation. After 250 training epochs, the model achieved consistent performance on the training set, demonstrating the efficiency of the proposed method.