Enhancing EDXRF Quantification Through Automated Matrix Determination

Abstract

Energy dispersive X-ray fluorescence spectrometry has been widely used for the analysis of trace and minor elements in applications that extend from biomedical to environmental assessment, due to its non-destructive nature, rapid analysis and suitable detection limits. However, EDXRF quantification is frequently hampered by matrix effects, introducing significant inaccuracies in the obtained results. In this work, we present an automated methodology for sample matrix determination using EDXRF spectra, leveraging on the analysis of the Compton and Rayleigh peaks of the characteristic lines of the X-ray tube anode. First, a model was created to fit these scattering peaks, allowing the plotting of a calibration curve that correlated the Compton-to-Rayleigh ratio with the average atomic number (Z) of the sample. Matrix composition was quantified using a developed algorithm combining support vector regression (SVR) and bootstrapping to optimize the determination of the best matrix composition. SVR with a Radial Basis Function kernel was applied to handle non-linear data, and Bootstrapping was utilized to train the algorithm, enhancing model generalization. The study demonstrates that the developed algorithm and matrix-based approach effectively quantified elemental compositions across various certified reference materials (CRMs). The chosen Matrix provided more accurate results, especially for heavier elements like Fe, Cu, and Zn, while deviations of around 20% were observed for lighter elements in biological matrices. In geological samples like phosphate rock and clay, heavier elements aligned well with reference values, but trace elements like Cu and Ni showed larger deviations due to low concentrations. Despite discrepancies for some elements like Pb in wood, the methodology proved effective, particularly for elements like Cr with minimal deviation, highlighting its versatility across diverse matrices. The methodology successfully integrated computational tools and open-source libraries to establish a reliable, efficient workflow for average atomic number determination and spectral analysis, achieving strong agreement with reference materials.