Improvement of Iterative Optimization Technology (Calibration-Free/ Minimum Approach) with Dimensionality Reduction of Spectra
Hiromasa Kaneko1, Koji Muteki2, Daniel O. Blackwood2, Yang A. Liu2 , Sonja S.Sekulic2, Kimito Funatsu1
1Department of Chemical System Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan
2Pfizer Worldwide Research & Development, ... [Show full abstract] Eastern Point Rd, Groton, Connecticut 06340, United States
Keywords: Process analytical technology, Iterative optimization technology, Dimensionality reduction
Process analytical technology plays an important role in the pharmaceutical industry. PAT is used extensively in process development, process understanding and process control. Often, quantitative measurements are desired/required and a calibrated model will have to be developed and implemented. The development, implementation and maintenance of these quantitative models are both resource and time intensive.
Muteki et al (2013)1. previously proposed calibration-free/ minimum approach, iterative optimization technology (IOT), which is used to predict the composition of a mixture while maintaining a similar predictability to calibration models (e.g. Partial Least Squares). It involves using only pure component spectra and mixture component spectra (without a calibration data set). Linear and non-linear IOT have been successfully applied (by offline/online) to some practical pharmaceutical process (e.g. blending, feed frame, solvent mixture, reaction, etc). However, a key (remaining) question was how to deal with a dependency (multi-collinear relationship) among pure component spectra during the optimization computation. For the mixture case which includes similar structured materials, it would be essentially difficult to provide good prediction on mixture component ratio. This problem has often been an important obstacle when applying IOT to mixture cases having more than several components.
This study presents a calibration method which can improve the prediction ability of IOT through reducing dimensionality of spectra with optimal selection of wavelength. It involves using Latent Variable Models (e.g. PCA, PLS) for dimensionality reduction of spectra in IOT and genetic-algorithm-based wavelength selection (GAWLS)2 for optimal wavelength selection. The proposed methods achieved higher predictive accuracy compared to the traditional IOT. The effectiveness of the proposed methods is demonstrated using simulation data where pure spectra have strong correlation and real industrial data of multicomponent mixtures.
* Muteki K, Blackwood DO, Maranzano B, Zhou Y, Liu YA, Leeman KR, Reid GL. Mixture Component Prediction Using Iterative Optimization Technology (Calibration-Free/Minimum Approach). Ind. Eng. Chem. Res. 2013;52:12258– 12268.
* Arakawa M, Yamashita Y, Funatsu K, Genetic algorithm-based wavelength selection method for spectral calibration. J. Chemom. 2011;25:10–19.