Application of unsupervised and supervised learning to a material attribute database of tablets produced at two different granulation scales

The purpose of this study is to demonstrate the usefulness of machine learning (ML) for analyzing a material attribute database from tablets produced at different granulation scales. High shear wet granulators (scale 30 g and 1000 g) were used and data were collected according to the design of experiments at different scales. In total, 38 different tablets were prepared, and the tensile strength (TS) and dissolution rate after 10 min (DS10) were measured. In addition, 15 material attributes (MAs) related to particle size distribution, bulk density, elasticity, plasticity, surface properties, and moisture content of granules were evaluated. By using unsupervised learning including principal component analysis and hierarchical cluster analysis, the regions of tablets produced at each scale were visualized. Subsequently, supervised learning with feature selection including partial least squares regression with variable importance in projection and elastic net were applied. The constructed models could predict the TS and DS10 from the MAs and the compression force with high accuracy (R2 = 0.777 and 0.748, respectively), independent of scale. In addition, important factors were successfully identified. ML can be used for better understanding of similarity/dissimilarity between scales, for constructing predictive models of critical quality attributes, and for determining critical factors.

Automated summary

The study demonstrates the usefulness of machine learning (ML) in analyzing a material attribute database from tablets produced at different granulation scales. Unsupervised learning is used to visualize the regions of tablets produced at each scale, while supervised learning with feature selection is applied to construct predictive models for critical quality attributes. The models can accurately predict the tensile strength (TS) and dissolution rate after 10 min (DS10) independent of scale, and identify important factors. Machine learning can be used to better understand similarity/dissimilarity between scales and determine critical factors.