Efficient input variable selection for calibration model design

In pharmaceutical processes, near-infrared spectroscopy (NIRS) is a key tool of process analytical technology (PAT), and very accurate calibration models need to be developed with NIR spectra. Partial least squares (PLS) regression, in particular, is accepted as a useful technique for calibration model design. When a calibration model is built, appropriate input variables have to be selected to achieve high estimation performance. Recently, a new methodology for selecting input variables based on nearest correlation spectral clustering (NCSC) has been proposed. Referred to as NCSC-based variable selection (NCSC-VS), it clusters input variables into some variable groups based on the correlation by using NCSC, and selects a few variable groups according to their contribution to output estimates. We report here an industrial application of NCSC-VS to calibration model design for a pharmaceutical process. NCSC-VS can select important variables and improve the estimation performance greatly in comparison to conventional variable selection methods.