Making isotopic composition artificially intelligent: Conceptual design and performance assessment of pattern recognition algorithms for uranium enrichment determination tasks using CZT and LaBr3(Ce) detectors

Abstract

There is currently a huge interest in the use of artificial intelligence (AI) technologies for various data analysis applications. Among these applications are radiation detection and spectra analysis tasks aiming at qualitative and quantitative information extraction on the assaying radioisotopes. AI tools are foreseen as promising techniques to deal with complex spectra analysis cases, where the traditional statistical tools are subject to deficiencies, due to high degree of noise, peak overlapping and statistical uncertainties because of the room temperature operation of detectors such as CdZnTe and LaBr 3(Ce). However, over the years there has been no comprehensive assessment of possible conceptual designs of such AI-based algorithms for spectra analysis purposes applied to uranium enrichment determination tasks. This paper analyzes different AI-based methodologies for the qualitative and quantitative analyses of uranium spectra and presents the AI-based Cluster Analysis via Machine Intelligence and Learning Algorithms (CAMILA) code for uranium enrichment determination. The performance assessment of multivariate and pattern recognition methodologies limits and possibilities is conducted in view of different physical conditions of the measurement system, such as the degree of attenuation and sample-to-detector distance, as well as impact of the neural network (NN) design on the algorithm performance. Tests are conducted on uranium spectra of CBNM certified standards with enrichment degrees from 0.31% up to 4.46% of 235 U atomic abundance measured with different statistical quality, as well as on simulated uranium spectra with a broad range of enrichment degrees. Implemented unfolding and analysis routines of the algorithms are described in detail and results are presented.