Artificial Neural Networks Applied in the Classification of Radionuclides from Gamma Spectroscopy and Computational Simulation

Abstract

The present study seeks to develop a classifier of radioactive sources based on gamma spectroscopy and artificial intelligence, which makes use of Keras and TensorFlow [, both free and open-source technologies. Through these technologies, an artificial neural network (ANN) was developed, which makes use of supervised machine learning and the backpropagation algorithm. The neural network was trained with a dataset of spectra from simulations performed by the Monte Carlo N-Particle 5 code (MCNP5). To achieve the objective of functioning as a classifier within the proposed scope, several versions of the neural network were evaluated, to study its performance, to determine important parameters such as a learning rate, and the optimizer used, among others, where RNA performance was evaluated by analyzing the network accuracy and loss curves. The generalization capacity was assessed by submitting spectra raised experimentally by an apparatus composed of a NaI(Tl) detector, a multichannel analyzer, and the Maestro software for the experimental data acquisition, which were used with the use of sealed radioactive sources of Co-60, Cs-137, and Eu-152. Six versions of the ANN were selected, capable of solving the proposed problem, with accuracy greater than 95%.