Development of machine learning algorithms for identification of radioisotopes using nuclear spectroscopy

• Grant number: 20/13879-4
• Support Opportunities: Research Grants - Innovative Research in Small Business - PIPE
• Start date: December 01, 2021
• End date: August 31, 2022
• Field of knowledge: Engineering - Nuclear Engineering - Applications of Radioisotopes
• Principal Investigator: Ubaldo Baños Rodríguez
• Grantee: Ubaldo Baños Rodríguez
• Company: Soluções e Instrumentações Avançadas Ltda
• CNAE: Pesquisa e desenvolvimento experimental em ciências físicas e naturais
• City: São Paulo
• Associated researchers: Daniel Alexandre Baptista Bonifacio ; Maurício Moralles
• Associated research grant(s): 23/00079-8 - Development of a radionuclide identifier based on nuclear spectroscopy and machine learning techniques, AP.PIPE
• Associated scholarship(s): 21/15175-7 - Design and implementation of an embedded signal acquisition system for radioisotope identification using machine learning, BP.TT; 21/14182-0 - Development of algorithms for analysis of acquisition system data for radioisotope identification using machine learning, BP.TT; 21/13679-8 - Development of machine learning algorithms for identification of radioisotopes using nuclear spectroscopy, BP.PIPE

Abstract

Situations such as radiological emergencies, decommissioning of nuclear facilities, border surveillance and cargo and sites monitoring requires a radioisotope identification system to ensure the nuclear safety of the country and prevent the irregular transportation of radioactive materials. The U.S. Homeland Security Department recommends a priority list of radioisotopes to be considered in these situations. The radioisotope identification is based on nuclear spectroscopy techniques, using detectors that record the time of arrival and peak of electrical pulses generated by the decay radiation. The events recording allows building a distribution of the peaks of electrical pulses according to their frequency of occurrence, which is converted into an energy spectrum for the study of incident radiation properties. Thus, spectral analysis allows the identification of radioisotopes and the determination of parameters such as activity and radionuclide emission probability. If the energy peaks of the spectrum are well defined, identifying them through processing algorithms is a relatively easy task. However, when there is an overlap of peaks or insufficient statistics, the algorithms tend to fail, and the process of radioisotope identification becomes a challenge. Such effects occur mainly in low-cost detectors, which have limited energy resolution and reduced detection efficiency, or when the detector's exposure time to the radioactive source is minimal. Due to the extensive and intensive use of ionizing radiation detectors, it is necessary to balance costs with the technical limitations of these systems. Thus, instead of using high-performance, high-cost semiconductor detectors, preference has been given to more economical, lower energy-resolution options, such as scintillator crystals coupled to silicon photomultipliers. However, lower energy resolution increases the probability of overlapping characteristic peaks, making it challenging to identify isotopes in heterogeneous samples by traditional nuclear spectra analysis algorithms. The objective of this project is to develop and evaluate machine learning algorithms for real-time identification of radioisotopes using nuclear spectroscopy techniques. The work proposes the use of Bayesian neural networks for the identification and quantification of radioisotopes in heterogeneous samples, with low-cost scintillation detectors, even in conditions that other algorithms fail, such as overlapping energy peaks or short acquisition time. Several neural network architectures will be evaluated concerning precision in isotope identification. For neural network training, emission spectra database of elements of interest will be simulated using the Geant4 radiation transport code. In this way, neural networks will be able to identify the elements present in a sample by analyzing the energy spectrum as well as determining their respective proportions in the material. (AU)