Influence of mixture design on multivariate prediction of PAHs in mixture spectra.

This work describes a simulation study aimed at establishing the impact of mixture design on the prediction ability of PLS regression models. Data sets are formed by multiplying UV absorbance spectra of 12 PAHs by their concentration profiles. In these case studies, either all possible mixtures of 1-12 components are used or randomly chosen selections of the mixtures. The effects of the number of samples and the number of concentration levels in the mixture designs on the results of the calibration are assessed. Comparisons are made between models formed using orthogonal fractional factorial mixture designs and those based on random designs. The applicability limits of the orthogonal designs are analysed in terms of actual concentration ranges of individual components in the mixtures.