(400g) Extended Iterative Optimzation Technology for Mixture Composition Estimation in Moving Powders

The calibration and maintenance effort for a near infra-red sensor and calibration based solution can be a time, labor and material-intensive exercise. A calibration-free method (iterative optimization technique or IOT) was recently published and applied onto mixture component prediction.[1]. IOT estimates fractions of individual components via optimization by rigorously imposing Beer-Lambertâ??s law and mixture constraints to the measured spectra, given the known spectrum for the pure materials. Advances in IOT theory propose the use of massive blind-search evolutionary algorithms to select spectral regions that can improve the performance of IOT in a real-time scenario.

The application of IOT onto powder flows is severely affected by scattering and uneven sample presentation (among others) which results in large deviations from Beer-Lambert in the mixture spectra. Other application-specific artifacts (like peak overlap) can also influence the performance of IOT, resulting in predictions with trends and biases that can range from mild, to completely unreasonable and inaccurate. This work presents an extension to IOT that preserves the calibration-free nature of the method while improving its real-time performance in the estimation of fractions of individual components in the powder stream of a continuous drug product manufacturing line.

Although still calibration free, EIOT requires at the very least measured spectra at a minimum of two levels of concentration for the active ingredient â?? this data set is readily available if a step change study is done in the line to determine mixing behavior of the system. The power of the method resides in the spectral differentiation across materials, as such, a data set consisting of spectral measurements at several levels of the individual components is much preferred [2].

EIOT fundamentally challenges the application of Beer-Lambert in relating the fractions of each material through the mixture spectra in the process and pure spectra collected in a static sample. As such, EIOT starts with the estimation of the dynamic pure component spectra that is mathematically necessary such that, the concentrations and the dynamic mixture spectra in the â??training setâ? is explained by Beerâ??s law. Multiple methods are presented to produce this estimate. It has been observed that the dynamic pure spectra estimated differ only in small features from the measured pure spectra. Once these pure spectra are estimated, the procedure continues as IOT would. These differences, although small play an important role in the performance of the method. Several examples are presented, contrasting the performance of a PLS model, IOT and EIOT in estimating the fractions of ingredients from spectra measured with a NIR instrument installed in the feed frame of a tablet press, at the end of a continuous drug product manufacturing line. The trends produced by EIOT lack undesired trends that are attributed to process effects onto the spectra, and represent a useful metric to monitor the real-time fraction of the materials in the feed-frame. Such an estimate can be used in combination with a deterministic model of the system to confirm mixture composition and confirm the state of control of the process.

[1] Muteki, K., Blackwood, D.O., Maranzano, B., Zhou, Y., Liu, Y.A., Leeman, K.R.and Reid, G.L. Industrial & Engineering Chemistry Research, 52, 12258 (2013).

[2] Eriksson, L.; Johansson, E.; Wikstrom, C. Mixture design--design generation, PLS analysis, and model usage. Chemometrics and Intelligent Laboratory Systems, 43, 1-2, 1-24. (1998)