Machine learning modelling of wet granulation scale-up using compressibility, compactibility and manufacturability parameters

Abstract

The purpose of this extensive study is to use a quality by design (QbD) approach and multiple machine learning algorithms in facilitating wet granulation process scale-up. This study investigated the extent of influence of both formulation and process variables. Furthermore, measured responses covered compressibility, compactibility and manufacturability of a powder blend. Finally, the models developed on laboratory scale samples were tested on pilot and commercial scale runs. Tablet detachment and ejection work were calculated from force-displacement measurements. Significant numerical and categorical input variables were identified by using a stepwise regression model and their importance evaluated by using a boosted trees model. Pilot scale runs resulted in the highest tablet tensile strength and compaction work as well as the highest detachment and ejection work. Critical quality attributes (CQAs) that were the most successfully predicted were the compaction, decompaction, and net work, as well as the tablet height. The most important input variable influencing all CQAs was the compaction force. Application of the boosted regression trees model resulted in the lowest Root Mean Square Error (RMSE) values for all of the responses. This work demonstrates reliability of predictions of developed models that can be successfully used as a part of a QbD approach for wet granulation scale-up.

Keywords

• quality by design
• artificial intelligence
• compaction work
• decompaction work
• elastic recovery