(T1530-06-35) Principal Component Analysis of Granule Size Distributions for Development of Continuous Twin-Screw Granulators

Purpose: Process understanding is a critical component of accelerating the implementation of continuous manufacturing (CM) systems. Twin-screw granulation (TSG) has shown promise for continuous wet granulation, but it requires an in-depth understanding of the effect of parameters on the granulated product. Various TSG parameters have been studied for their effect on the granule size distribution (GSD). Overwhelmingly, liquid solid ratio (L/S) has been shown to significantly affect the GSD by promoting granule growth.¹ Currently, d-values (d10, d50, dv50, etc.), which are percentiles within the GSD, are employed as the primary data analysis for GSDs. Additionally, d-values have demonstrated success as potential indicators of process performance.² In the context of process development, however, d-values may not fully describe the changes in the GSD as a result of process parameter changes. These distributions are often complex, non-normal distributions, which can create problems when relying on d-values to fully describe the effect of process parameter changes.

In this work, the ability of PCA to differentiate GSDs was investigated. Simulated distributions were generated to demonstrate the lack of information provided by d-values. A time-dependency study was performed to assess the ability of PCA to track trends in GSDs over time. Finally, similar GSDs were compared via PCA to evaluate methods to characterize the true difference between distributions. Principal component score plots are used to demonstrate the separation of GSDs based on distribution shape, and the loading plots are used to highlight the meaningful changes in variance structure of the GSDs.

Methods: The Dome-Ex Granulator and DG-Dryer system (Dalton, Osaka, Japan) was used to granulate a model theophylline formulation. Two experiments were used to examine the ability of PCA to extract differences between GSDs and provide meaningful results in the context of CM method development. Granules were collected at each experimental condition and analyzed with dynamic image analysis via SolidSizer (J.M. Canty, NY, US). The resulting granule sizes were used to construct GSDs representing each condition, followed by PCA in MATLAB (Mathworks Inc., MA, US). The number of principal components included was selected based on the total variance captured, often resulting in the selection of two or three principal components. Additionally, MATLAB was used to calculate the d-values for each GSD, and this data was compared with PCA scores plots to determine the ability of d-values to represent changes to the GSD. For this work, ‘fines’ refers to particles smaller than 300um.

Results: The first investigation examined two common GSD types to demonstrate the lack of information provided by d-values. Figure 1 shows a bimodal distribution and a normal distribution with the d-values marked with dashed lines. The lack of information given by d-values can affect the interpretation of processing conditions and their effects on the granulated product. The next investigation was the analysis of time-dependent GSDs from a TSG system. One set of parameters was kept constant for the duration of the experiment to assess any experimental drift over time. The analysis by d-values and PCA. PCA created disparate interpretations of the resulting material. PCA found no significant difference between the granule size distributions throughout the experiment. For reference, a GSD for raw material was included to demonstrate the ability of PCA to separate the distributions. The scores and loadings plots are shown in figure 2a and 2b, respectively. The dv10, however, indicated drift throughout the experiment, which is not consistent with the fines content as seen in figure 2c, particularly between 50 and 60 minutes. There is an extreme drop in dv10, but the fines content stays relatively constant. This is indicative of the limited, if not misleading, information provided by the d-values. Finally, similar experimental conditions were used to investigate the ability of PCA to differentiate similar distributions. The results are shown in figure 3. The separation seen in figure 3a demonstrates ability of PCA to extract true differences between the GSDs, and the loadings indicate the difference to be in the fines and the oversize bins. Comparing the PCA results with d-values reveals that dv50 is potentially applicable for determining the presence of a difference between the distributions, but it does not appear to fully describe the differences in distribution shapes.

Conclusion: First, the lack of information provided by d-values is demonstrated. The time-dependency study shows trending in dv10 while the true distributions are not significantly different from one another, indicating PCA’s ability to analyze the shape of the distribution rather than a percentile descriptor. Finally, similar distributions are compared to demonstrate PCA’s ability to differentiate GSDs and provide meaningful information via loadings plots regarding the differences between them. Overall, PCA is shown as a valuable data analysis method for extracting differences between GSDs, which contributes greatly to process understanding and can ultimately facilitate the implementation of CM. Future investigations will assess the ability of PCA to describe the effects of various processing conditions.

References: 1. Dhenge, R. M., Fyles, R. S., Cartwright, J. J., Doughty, D. G., Hounslow, M. J., & Salman, A. D. (2010). Twin screw wet granulation: Granule properties. Chemical Engineering Journal, 164(2-3), 322-329.
2. El Hagrasy, A. S., Cruise, P., Jones, I., & Litster, J. D. (2013). In-line size monitoring of a twin-screw granulation process using high-speed imaging. Journal of Pharmaceutical Innovation, 8(2), 90-98.

Acknowledgements: The authors wish to gratefully acknowledge the support from LCI for this work.


Comparison between normal (orange) and bimodal (blue) distributions with identical d-values (dashed lines).


A-PC scores plot showing the separation of granulated product from raw material as well as the consistency of product over time with a 95% confidence ellipse. B-PC1 loading showing the primary difference between the GSDs to be fines content. C-Comparison of dv10 and fines content, ultimately showing the inability of dv10 to capture the true fines content.


A-Shows the ability of PCA to separate similar GSDs. B-PC1 loading plot showing the influence of both fines and oversize particles on the separation of GSDs. C-Comparison of d-values for two similar conditions where dv10 and dv90 overlap, but dv50 shows separation between the conditions.