Predicting pharmaceutical powder flow from microscopy images using deep learning

The powder flowability of active pharmaceutical ingredients and excipients is a key parameter in the manufacturing of solid dosage forms used to inform the choice of tabletting methods. Direct compression is the favoured tabletting method; however, it is only suitable for materials that do not show cohesive behaviour. For materials that are cohesive, processing methods before tabletting, such as granulation, are required. Flowability measurements require large quantities of materials, significant time and human investments and repeat testing due to a lack of reproducible results when taking experimental measurements. This process is particularly challenging during the early-stage development of a new formulation when the amount of material is limited.

To overcome these challenges, we present the use of deep learning methods to predict powder flow from images of pharmaceutical materials. We achieve 98.9% validation accuracy using images which by eye are impossible to extract meaningful particle or flowability information from. Using this approach, the need for experimental powder flow characterization is reduced as our models rely on images which are routinely captured as part of the powder size and shape characterization process. Using the imaging method recorded in this work, images can be captured with only 500 mg of material in just 1 hour.

This completely removes the additional 30 g of material and extra measurement time needed to carry out repeat testing for traditional flowability measurements. This data-driven approach can be better applied to early-stage drug development which is by nature a highly iterative process. By reducing the material demand and measurement times, new pharmaceutical products can be developed faster with less material, reducing the costs, limiting material waste and hence resulting in a more efficient, sustainable manufacturing process. This work aims to improve decision-making for manufacturing route selection, achieving the key goal for digital design of being able to better predict properties while minimizing the amount of material required and time to inform process selection during early-stage development.

Download the research paper as PDF:Predicting pharmaceutical powder flow from microscopy images using deep learning

or continue reading here on predicting pharmaceutical powder flow


List of materials

4-Aminobenzoic acidSigma-Aldrich
Cetyl alcoholSigma-Aldrich
Avicel PH-101Dupont
Azelaic acidSigma-Aldrich
Benecel K100MDupont
Calcium carbonateSigma-Aldrich
Calcium phosphate dibasicSigma-Aldrich
Cholic acidSigma-Aldrich
Dimethyl fumarateSigma-Aldrich
FastFlo 316Dupont
Granulac 230Meggle Pharma
Ibuprofen 50BASF
Ibuprofen 70Sigma-Aldrich
Lubritose mannitolKerry
Lubritose MCCKerry
Lubritose PBKerry
Magnesium stearateRoquette
Magnesium stearateSigma-Aldrich
Mefenamic acidSigma-Aldrich
Methocel DC2Colorcon
Microcel MC-200Roquette
Mowiol 18-88Sigma-Aldrich
Paracetamol granular specialSigma-Aldrich
Paracetamol powderSigma-Aldrich
Parteck 50Sigma-Aldrich
Pearlitol 100SDRoquette
Pearlitol 200SDRoquette
Pluronic F-127Sigma-Aldrich
Potassium chlorideSigma-Aldrich
Sodium stearyl fumarateSigma-Aldrich
Span 60Sigma-Aldrich
Stearyl alcoholSigma-Aldrich

Matthew R. Wilkinson, Laura Pereira Diaz, Antony D. Vassileiou,  John A. Armstrong,  Cameron J. Brown, Bernardo Castro-Dominguez and  Alastair J. Florence

DOI: 10.1039/D2DD00123C , Digital Discovery, 2023, Advance Article

You might also like