Variable selection and multivariate calibration models for X-ray fluorescence spectrometry

Abstract

Application of partial least squares models to quantitative XRF analysis is shown to be dependent on the efficiency of the variable selection routine used to choose the spectroscopic variables employed in the calibration model. Three variable selection techniques have been examined, based on cumulative covariance of emission intensity with analyte concentration, signal-to-noise ratio of linear sensitivity of individual variables to error variance and signal-to-noise ratio of logarithmic sensitivity of individual variables to error variance. Applying the models to a wide variety of analytes in a range of sample types, logarithmic scaling of the sensitivity term in the signal-to-noise ratio is demonstrated as providing sensible and efficient selection of appropriate variables. Results obtained are comparable to those using element selective correction models such as Lucas–Tooth and Price, which require user knowledge of the sample and analytical conditions.