Improvement of PLS regression-based XRF spectroscopy quantification using multiple step procedure and Monte Carlo simulation

Partial least squares (PLS) regression-based methods have been proven to be a good alternative for quantification in X-ray fluorescence spectroscopy. These methods are fast and easy to use though giving satisfactory results under certain conditions. One of these conditions is the necessity of having a great number of spectra to build the model (training set). The choice of the constituent concentration range in the training set has a big influence on the accuracy of the model. Better accuracy is obtained if the model is built in relatively narrow regions containing (or close to) the real concentration value. In the present work, Monte Carlo (MC) simulated spectra are used to form the training set. The advantage to use MC generated training spectra is the unlimited availability of perfect standards. This paper aims to improve the accuracy of the method by introducing a multiple step procedure in order to build the PLS model using narrow concentration range close to (or containing) the real concentration values in the samples to be measured. This approach consists of an initial guess of the constituents' concentrations and a preliminary PLS model before building the final model. The prediction of ten MC simulated alloy standard samples containing Ti, Mn, Fe, Co, Cu, Zn, Sr, Zr, and Mo using this method allowed to have average relative prediction errors less than 5% for elements with narrow concentration ranges. Copyright © 2013 John Wiley & Sons, Ltd.