Machine learning approaches to the prediction of powder flow behaviour of pharmaceutical materials from physical properties

Abstract

Understanding powder flow in the pharmaceutical industry facilitates the development of robust production routes and effective manufacturing processes. In pharmaceutical manufacturing, machine learning (ML) models have the potential to enable rapid decision-making and minimise the time and material required to develop robust processes. This work focused on using ML models to predict the powder flow behaviour for routine, widely available pharmaceutical materials. A library of 112 pharmaceutical powders comprising a range of particle size and shape distributions, bulk densities, and flow function coefficients was developed. ML models to predict flow properties were trained on the physical properties of the pharmaceutical powders (size, shape, and bulk density) and assessed. The data were sampled using 10-fold cross-validation to evaluate the performance of the models with additional experimental data used to validate the model performance with the best performing models achieving a performance of over 80%. Important variables were analysed using SHAP values and found to include particle size distribution D10, D50, and aspect ratio D10. The very promising results presented here could pave the way toward a rapid digital screening tool that can reduce pharmaceutical manufacturing costs.