NIR Calibration Chemometrics

NIR analysts often use a statistical methodology called chemometrics to calibrate an NIR analysis. Chemometrics is a specialized branch of mathematical analysis that uses statistical algorithms to predict the identity and concentration of materials. Chemometrics is heavily used in NIR spectral analysis to provide quantitative and qualitative information about a variety of pure substances and mixtures. NIR spectra are often the result of complex, convoluted, and even unknown interactions of the different molecules and their environment. As a result, it is difficult to pick out a spectral peak or set of peaks that behave linearly with concentration or are definitive identifiable markers of particular chemical structures. Chemometrics uses statistical algorithms to pick out complex relationships between a set of spectra and the material s composition and then uses the relationship to predict the composition of new materials. Essentially, the NIR system, computer, and associated software are trained to relate spectral variation to identity and then apply that training to new examples of the material. 

There are a variety of chemometric models for both qualitative and quantitative analysis. A common quantitative model is called partial least squares (PLSs) (or projection to latent structure). PLS condenses the thousands of individual data points across the whole spectrum in all of the standards to a few sources of variation that contrast broad bands of frequencies. These condensed sources of variation are called factors and are selected as part of the model in order of importance in accounting for the variation. The first factor of a good model will account for a large part of the variation and subsequent factors will account for the remaining variation. In practice, only a limited number of factors are required to account for most of the variation. 

A practical example of the chemometrics approach is described in the reference by Heil. 

In order to develop a PLS calibration model for moisture, oil, protein, and linolenic acid in soybeans, a total of 207 calibration and 50 validation samples were used to build and validate the calibration model described in this chapter. While this may seem like a lot of work, 64 coaveraged scans were collected using a Thermo Scientific Antaris IIFTNIR instrument in 30 s with a resolution of 8 cm k No processing or grinding of the raw soybeans was required, and once the method is validated, no additional calibration is required. The wet chemical methods to measure these four parameters would require grinding, digestion, extraction, and multiple analytical techniques. 

---