Multicomponent analysis of vapor phase infrared spectra using multivariate regression methods

The concepts of multivariate regression methods, partial least-squares (PLS), principal component regression (PCR), classical least-squares (CLS), and inverse least-squares (ILS), are described. The performance of these methods applied to vapor-phase Fourier transform mid-infrared spectra of mixture of the xylene isomers and five chlorinated hydrocarbons, measured at 4 cm{dollar}sp{lcub}-1{rcub}{dollar} and 8 cm{dollar}sp{lcub}-1{rcub}{dollar} resolution, respectively, is compared. The calibration models developed from these data are used to predict the concentration of unknown samples that are not included in the calibration set. Much better results were obtained on these spectral data with the factor-based regression methods, PLS and PCR, than with CLS or ILS, with PLS being faster than PCR in modeling a large calibration set. The calculated standard error of calibration (SEC) and standard error of prediction (SEP) were used to evaluate the prediction capability of regression models for calibration and validation samples, respectively.; Utilization of PLS for quantification of vapor-phase spectra measured at low spectral resolution was investigated. Vapor-phase infrared spectra of five chlorinated hydrocarbons, measured at different resolutions of 1 to 64 cm{dollar}sp{lcub}-1{rcub}{dollar}, were obtained from their corresponding interferograms truncated at different lengths. The truncated interferograms were then multiplied by different apodization functions prior to Fourier transformation. The effects of apodization function, noise levels, measurement time, and the spectral resolutions concerned on the quantitative analysis of these spectra were investigated. The more accurate PLS results were obtained from the spectra measured at low resolution, 16 cm{dollar}sp{lcub}-1{rcub}{dollar}, with 16 scans.; The quantification error due to the difference in temperature between the spectral data used for calibration and for validation was determined. Quantitative errors in PLS regression due to temperature fluctuations in sample set were minimized by developing a model from a calibration set containing 124 samples measured over a temperature range of 30{dollar}spcirc{dollar}C.; Finally, the effect of both temperature and resolution on library searching spectral data obtained when a gas chromatography is coupled with a Fourier transform infrared spectrometer experiment is studied by analysis of variance. The better results were obtained in compound identification when the experimental parameters for the sample spectra of interest and reference spectra in the library were closely matched.