Nuclear Forensic Analysis via Machine Learning Assisted Laser-based Spectroscopy and Spectral Imaging

Abstract

Nuclear forensics (NF) is a systematic and scientific methodology designed to identify, categorize, and characterize seized nuclear and radiological materials (NRM). The aim of NF is to reveal the geographical origin, process/production history, age and intended use of the NRM to prevent future diversions and thefts, thereby strengthening the national security of a country. The complexity of the signatures utilizing the existing methods poses an analytical and interpretation challenge. Hence, the need to develop rapid, non-invasive and non-destructive techniques to speed up NF investigations. Laser Based Induced Breakdown Spectroscopy (LIBS) fingerprints the elements associated with the spectral peaks, while Laser Raman microspectrometry (LRM) uniquely identifies specific chemical compounds and microstructures in a sample based on molecular vibrations. Although these methods have high accuracy and versatility following little or no sample preparation, their practical utility is limited due to the complexity of the samples and the interpretative challenges of multivariate data. Machine learning (ML) techniques can overcome these limitations and help analyse this complicated and bulky data. LIBS and LRM combined with ML possess the power to conduct direct, rapid NF analysis of limited size NRM with accuracy and precision. The uranium lines at 386.592 nm, 385.957 nm and 385.464 nm were identified as NF signatures of uranium in cellulose and uranium ore surrogates (uranium mineral ores and high background soil samples). The detection limit for uranium in cellulose was determined at 76 ppm. Multivariate calibration models in artificial neural network (back-propagation algorithm) were developed using resonant and weak uranium lines. The calibration model using weak U-lines predicted the uranium concentration in the certified reference material (CRM) RGUMix (101 ppm) and RGU-1 (400 ppm) at relative error of prediction (REP) = 2.97% and 2.25% respectively, while using resonant U-lines at REP = 69.31% and 4.25%, respectively. The calibration model utilizing weak U-lines predicted the uranium content in the uranium mineral ores in the range

of (112 - 1000) ppm. Application of principal component analysis (PCA) on the complete LIBS spectra of uranium ore surrogates revealed patterns that were related to their origin. PCA applied to selective spectral regions of uranium mineral ores successfully grouped them into their mineral mines (origin). NF signatures associated with uranium molecules in uranyl nitrate, uranyl sulphate, uranyl chloride and uranium trioxide samples were identified at 865 cm-1, 868 cm-1, 861 cm-1, and 848 cm-1respectively using LRM (laser λ= 532 nm, 785 nm). Spectral imaging on simulate samples of uranium and uranium ore surrogates using NF signatures demonstrated the distribution of uranium molecules. Thus, ML assisted laser-based spectroscopy and spectral imaging, have the potential to not only perform rapid, direct and minimally intrusive qualitative and quantitative analysis of trace uranium, but also aid in the source attribution of uranium ore surrogates and distribution of uranium molecules.