Search

Multiway methods are tested for their ability to explore and model near-infrared (NIR) spectra from a pharmaceutical batch process. The study reveals that blocking of data having a nonlinear behaviour into higher-order array can improve the predictive ability. The variation in each control point is independently modelled and N-way techniques overcome the nonlinearity problem. Important issues as variable selection and how to fill in for missing values have been discussed. Variable selection was shown to be essential to be able to perform multiway modelling. For spectra not yet monitored, use of mean spectra from calibration set gave close to the best results. Decomposing the spectra by N-way techniques gave additional information about the chemical system. To support the results simulated data sets were used.

Most spectra with structural information are smooth. Making use of this feature, a new procedure is proposed for automatic chemical rank determination. In order to avoid accumulation of noise, key spectra instead of the full matrix are analyzed in the procedure. A simplified morphological factor, morphological score (MS), is proposed. The noise level of morphological analysis is established with the help of F -test. The proposed procedure was successfully applied in the study of self-association behavior of alcohol. Results show that two classes of aggregates contribute to near infrared (NIR) spectra besides the monomer. Data sets produced by other techniques were also tested.