[303795] Artigo – CBMRI – Congresso Brasileiro de Metrologia das Radiações Ionizantes

Resumo: The main objective of the current study is to develop a neural network for the prediction of response matrix for neutron spectrometry with a Bonner sphere. The Bonner multisphere spectrometry system, more commonly known as Bonner Spectrometer (BS), is a system that consists of a set of moderator spheres, where in the center of each sphere it is possible to accommodate a thermal neutron detector.

However, in this system each sphere behaves as a different detector, therefore, obtaining information from a neutron spectrum through the BS requires knowledge of the response of each sphere as a function of the neutron energy. This process introduces na imperative to obtain the response of a set of spheres through the response function that must be determined in order to characterize the neutrons that pass through the system. This response function, which allows access to the fluence values of the neutrons in the various energy ranges, is in practice replaced by a response matrix. Obtaining this response matrix is not trivial and involves a complex calculation process that includes Monte Carlo simulation from experimental data. The present work seeks to circumvent this difficulty using machine learning techniques with neural networks. The Laboratory of Neutron Metrology (NL/NLMIR) of the National Laboratory of Metrology of Ionizing Radiation (NLMIR) of the Institute of Radioprotection and Dosimetry of Brazil (IRD), which dominates the Bonner spectrometry technique where the development of the present work took place, has given particular attention to techniques for determining the response matrix, in particular the use of neural networks given the data set and experience it has accumulated with its Bonner sphere over the years. An MLP (Multi Layer Perceptron) network was developed in this work to obtain the response matrix. This network is tested by changing its different parameters and comparing its performances with other neural networks. The observed performances demonstrate that the developed technique serves as a solid alternative for obtaining the response matrix in neutron spectrometry.