(725b) Material Sparing Approaches for Predicting Powder Flow Using Machine Learning Methods

The performance of pharmaceutical powders is often determined by the distributions of both shape and size of the constituent particles, but typically they are described by simplistic, and often inappropriate descriptors such as D₅₀. Consequently, most of the information about the materials being characterized is lost and/or misrepresented. Image analysis enables the imaging of all particles within a sample and measurement of multiple particle characteristics resulting in an information rich data set. The complex interdependency between particle size and shape and its combined effect on powder performance metrics such as flowability is a prime problem to be tackled using machine learning techniques.

In this work, we explore the application of unsupervised and supervised machine learning techniques to predict the effect of particle morphology of pharmaceutical powders on their flow function coefficient (FFC). To achieve this, we develop novel feature extraction methods using principal component analysis and convolutional neural networks that capture the particle size and shape distribution of powders and lay foundations for a machine learning framework for aiding material selection in pharmaceutical development. Our results indicate that it is possible to distinguish non-flowing powders (FFC <= 2) from other flow classes (FFC > 2) with high confidence, thereby providing an early indication of poor flow and possible recourse using this material sparing approach. The broad applicability of this approach for predicting other powder performance metrics as well as applicability in other areas of drug development will also be discussed.