Decoding the interfering L-lines by artificial neural network-based modeling for direct analysis of lanthanides in water samples using Total reflection X-ray fluorescence spectrometry

Abstract

In this study, an Artificial Neural Network (ANN)-based methodology for the direct analysis of lanthanides using their highly interfering L-lines as characteristic signatures has been demonstrated. The model was validated on samples containing ten lanthanides (La, Nd, Sm, Eu, Gd, Tb, Dy, Ho, Tm, and Lu) at varying concentrations ranging from 0.25 microg/mL to 5.25 microg/mL. A total of 20 samples were prepared, with each sample analyzed in five replicates using Total Reflection X-Ray Fluorescence (TXRF). The TXRF data were first analyzed using a classical peak-fitting methodology, yielding a relative error of 14.4% and a precision (RSD) of 9.5%. The same dataset was then used to train an optimized ANN model, which significantly improved the analytical parameters, achieving a relative error of 8.5% and precision of 1.9%. Further validation was performed using reverse osmosis (RO) drinking water samples spiked with lanthanides, where the ANN model demonstrated a relative error of 10.1% and precision of 1.4%. This comparative study clearly highlights the superiority of the ANN-based methodology over traditional peak-fitting approaches, demonstrating its potential for more accurate and precise quantification of lanthanides in complex spectral environments.