Use of CLS to Understand PLS IR Calibration for Trace Detection of Organic Molecules in Water

Abstract

Partial least-squares (PLS) and principal component regression (PCR) methods applied to spectral data can generally provide excellent quantitative analysis precision, but extraction of qualitative spectral interpretation from the models can be more difficult. For example, we have achieved sensitivity in the parts-per-million range for polar organic compounds in aqueous solutions using infrared (IR) spectroscopy and modified sol-gel-coated ATR sensors. The interpretation of PLS or PCR loading vectors obtained from calibrations involving orthogonally designed solutions of acetone and isopropanol in water yields a misleading understanding of the mechanism for the IR detection of isopropanol on these sensors. Examination of the loading vectors from PLS or PCR or the first weight-loading vector from the PLS model would suggest that the spectral calibration is based largely on the interaction of the isopropanol with the surface modifier of the sol-gel coating. However, a classical least-squares (CLS) analysis of the data shows clearly that this interaction is not a significant source of the spectral calibration, but rather the calibration is primarily due to the spectroscopic signal of the isopropanol analyte. In this case, the misleading qualitative interpretation of the PLS and PCR models is the result of the spectral variation being dominated by the effects of spectrometer drift. CLS can overcome this problem if a parameter is included in the CLS calibration that adequately represents the drift. In the example presented here, time of spectral data collection is an appropriate drift-related parameter that can be added to the CLS calibration concentration model in order to provide the qualitative information needed to correctly interpret the spectral data. Other methods to include the effects of spectrometer drift in the CLS model are also presented.