Raman Spectroscopy is capable of differentiating similar molecules with different crystal lattice structures given the lattice vibrations of the entire molecule within the lattice structure of the crystal is different.  In addition, the use of fiber optics to collect data through an immersion probe allows analysis of solid phase composition in real-time.  However, the Raman intensity of the solids depends on the amount of inelastic scattering of the solids detected by the analyzer within the detection zone.  As a result, the relative Raman intensity corresponding to the diastereomers in a slurry will be impacted by a number of solid-state factors.  It has been suggested that Raman intensity with respect to different polymorphs may be a function of particle size and shape.  This is based on the assumption that Raman signals primarily come from the surface of the crystals.  Additionally, slurry density may be another solid-state factor since the number of crystals inside the detection zone will influence the Raman intensity of the solids.  In theory, the Raman spectrum will be affected by the amount of solvent and solids detected.  Thus, slurry density should impact the Raman signal intensity of the solid phase.

In the present work, the first objective was to examine whether information provided by Raman spectroscopy is sufficient or whether it needs to be complemented by additional process measurements in order to provide an accurate estimation, through a Partial Least Square (PLS) model, of the solid composition of one of the two diastereomers involved in the production of an active pharmaceutical ingredient, denoted here as compound A.  The selection of process variables was based on the cooling crystallization procedure of compound A.  Since the changing temperature, slurry density, and percent composition of the diastereomers would affect peak position and intensity, those were the variables selected in our modeling task.  Partial Least Square regression (PLS) was used to quantify the composition of the diastereomers mixture.  The second objective was to examine whether additional subsets of the calibration data needed to build models that better represent data variations.  Principle Component Analysis (PCA) was used to examine the data and separate data into two different subsets according to slurry density.   

This presentation addresses the estimation of fractional composition of two diastereomers during crystallization.  The estimation is obtained through a Partial Least Square (PLS) model that utilizes on-line Raman spectroscopy and additional process information such as temperature and slurry density.  12 PLS models were constructed with the same 95 calibration standards using Raman spectra, temperature and/or slurry density data.  They differ from each other on whether they model all the data or one of two subsets and on whether they involve temperature and/or slurry density along with the spectral data.  The models were further tested and compared against data from four 250mL scaled crystallization experiments.  It was shown that in-situ Raman spectroscopy is capable of differentiating diastereomers in a crystallization slurry, provided the changing process parameters of temperature and slurry density are included in the model.  Models developed for the high or low slurry density subsets of the data were more accurate than the corresponding models developed for the whole data.  PCA analysis was used for the effective separation of the training data into two subgroups.  The sample-to-model distance also proved useful in selecting the PLS model to be used with a new data point to estimate the diastereomer composition.