Comparing Deep Learning Architectures On Gamma-Spectroscopy Analysis For Nuclear Waste Characterization

Neural networks, particularly deep neural networks, are used nowadays with great success in several tasks, such as image classification, image segmentation, translation, text to speech, speech to text, achieving super-human performance. In this study, we explore the capabilities of deep learning on a new field: gamma-spectroscopy analysis, comparing the classification performance of different deep neural network architectures. We choose VGG-16, VGG-19, Xception, ResNet, InceptionV3, and MobileNet architectures, which are available through the Keras Deep Learning framework to identify several different radionuclides (Am-241, Ba133, Cd-109, Co-60, Cs-137, Eu-152, Mn-54, Na-24, and Pb-210). Using an HPGe detector to acquire several gamma spectra from different sealed sources to create a dataset used for the training and validation of the comparison of the neural network. This study demonstrates the strengths and weaknesses of applying deep learning on gamma-spectroscopy analysis for nuclear waste characterization.