Design of a Radioisotope-excited Edxrf System for Rare Earth Elements Analysis in Geological Samples

Abstract

Quantification of rare earth elements (REEs) using Energy Dispersive X-ray Fluorescence (EDXRF) is severely affected by presence of other elements, concentration, nature and energy of the excitation source, detection system, and analysis technique. Conventional REE analysis methods such as NAA, XRF, ICP-(OES and MS) are expensive, unavailable, and involve lengthy sample preparation. Robust elemental quantification using EDXRF and machine learning techniques have been demonstrated in many settings. Therefore, this study aimed at designing a radioisotope excited EDXRF instrument and using chemometrics and machine learning (ML) to quantify REEs in geological materials and starch. The instrument was built using an annular Americium-241 excitation source with an activity of 106 mCi and a peltier cooled SDD Detector. Analytical samples were prepared by schematically mixing REEs salts; Dy, Y, and Ce in geological and starch matrices. The EDXRF setup was used to acquire spectra and R software was used for data visualization, feature selection, scatter ratio correction, performance of PCA for dimension reduction, and to build ML models; SVR, ANN, and RF. Instrument shielding resulted in reduction of doses from 1.68 mSv/h with sample chamber door open to 250 nSv/h while closed. Results of scatter ratio correction established that regions for rock and starch matrices were different, 16.4~17.4 keV for rock and 18.7~20.6 keV for starch. RF model of Cerium in rock attained lowest root mean squared error of prediction (RMSEP) of 106 ppm at 57% accuracy using 9 PCs with limit of detection (LoD) of 9 ppm. RF model of Dy in rock attained the lowest RMSEP of 79 ppm at an accuracy of 41% using 3 PCs with LoD of 20 ppm. RF model of Y in rock attained lowest RMSEP of 140 ppm at an accuracy of 99.9% using 2 PCs with LoD of 64 ppm. RF model of Ce in starch attained lowest RMSEP of 30 ppm at 90% accuracy using 7 PCs with LoD of 6 ppm. NN model of Dy in starch attained lowest RMSEP of 25 ppm at 95% accuracy using 5 PCs with LoD of 7 ppm. NN model of Y in starch attained lowest RMSEP of 112 ppm at an accuracy of 99.99% using only 1 PC with LoD of 71 ppm. RF model of Ti in starch attained lowest RMSEP of 41 ppm at 78% accuracy using 5 PCs with LoD of 7 ppm. NN model of Nb in starch attained lowest RMSEP of 14 ppm at 98% accuracy using 9 PCs with LoD of 4 ppm. This study showed that with limited resources, an XRF instrument setup can be used with machine learning techniques to quantify REEs in geological and other matrices.