The development and application of online modelling methods

Chemometrics and Design of Experiments (DOE) are fast becoming integral parts of process analysis and incorporated into the resulting advances in technology. To this end, two major studies were undertaken to explore the existing methods of modelling using both traditional and modern forms of process analytical technology, and to create new methods using the most current developments in the field.The first study involved the use of chemometrics and DOE with low-resolution NMR FID spectra of a series of polymers that were collected over a period of ten months. Accompanying the NMR FID spectra were the associated laboratory reference measurements for a series of quality assurance parameters. This information was used to build an online prediction model for the Xylene Soluble (XS) content of polymer pellets. The installation of the online model was accomplished in numerous stages during which various sample selection methods, including work by Shenk and Westerhaus, were developed and evaluated. The intrinsic nature of the NMR data meant that traditional methods of sample selection could not be employed. The final model used the principal component analysis scores as a means of selecting samples for calibration. DOE was used to determine the best method of pre-processing to be applied to the data prior to partial least squares modelling. The final PLS model was evaluated and the error in prediction for the XS content was found to be 0.672%. The success of this project lead to the installation of this product online at the point of analysis in December 2006.The second study employed chemometrics and DOE with a more traditional method of process analytical technology, the NIR spectral analysis of pharmaceutical tablets. The NIR spectra of over 250 tablets were collected over three production campaigns from 1997 to 1999. Accompanying the NIR spectral data were the chemical and physical tablet parameters, active pharmaceutical ingredient, weight, and tablet thickness. The sample selection techniques developed as part of the polymer study were evaluated. In order to correct for variations due to specular and diffuse scattering effects, extended multiplicative scatter correction was applied to the data. As with the polymer study, DOE was used to determine the best method of data pre-processing prior to the partial least squares modelling. The best method of sample selection for this study was found to be the use of the condition number. The final prediction models for the active pharmaceutical ingredient, weight, and tablet thickness were produced. The final step for this study would be to apply this model online at the point of analysis in the same manner as the polymer study