(103d) Prediction of Powder Flow of Pharmaceutical Materials from Physical Particle Properties Using Machine Learning

Understanding powder flow and how it affects pharmaceutical manufacturing process performance remains a significant challenge for industry, adding cost and time to the development of robust production routes. It is therefore crucial improve our understanding of powder flow and how this relates in particular to both tableting and capsule filling. Tablet formation can be addressed by several techniques; however, direct compression (DC) and wet granulation (WG) are the most widely used in industrial operations. DC offers lower-cost manufacturing as a consequence of the fact that it provides streamlined process with fewer steps compared with other unit operations. However, to achieve the full potential benefits of DC for tablet manufacture, this places strict demands on material flow properties, blend uniformity, compactability, and lubrication, that need to be satisfied. On the other hand, WG improves flowability and compactability while preventing segregation through the use of binding agents and secondary process steps although it is not suitable for materials which are sensitive to heat or moisture. WG is, therefore, more expensive and time consuming. Due to these economic advantages, DC is increasingly the preferred technique for pharmaceutical companies for oral solid dose manufacture, consequently making the flow prediction of pharmaceutical materials of increasing importance ^[1,2]. Flow properties are influenced by the particle size and shape, which are defined during crystallization and/or milling processes. Currently suitability of raw materials and/or formulated blends for DC requires detailed characterization of the bulk properties. A key goal of digital design within Industry 4.0 is to be able to better predict properties whilst minimizing the amount of material required and time to inform process selection during early stage development.

This work aims to improve decision making for manufacturing route selection in the early stages of the drug product manufacturing process development when time and amount and cost of available material are at a premium. A Machine Learning (ML) model approach is proposed to predict the flow properties, specifically the flow function coefficient (FFc) of a new material from its physical particle properties. The aim is that the model produced as an outcome can help decide whether a new material is suited for DC.

Methods

To build a ML model, and hence, to make reliable predictions, a large data set is necessary. Particle size, shape distribution, and powder FFc of more than 100 materials were collected as training data to the supervised and unsupervised ML algorithms used. Particle size and shape distribution were determined using dynamic image analysis (QIC/PIC – Sympatec). This equipment is appropriate for materials in a range of sizes between 30 µm to 8,665 µm, and it only requires less than 2 grams. A Freeman FT4 (Freeman Technology, Malvern, UK) was used to carry out a shear cell test to measure the FFc. The consolidation stress and pre-shearing normal stress in the experiments was 9 kPa. Normal stress for shearing at 7 kPa, 6 kPa, 5 kPa, 4 kPa, and 3 kPa were used, and the sample is sheared to obtain five yield points at different normal and shear stress. Triplicates of each sample were measured using the 25 mm x 10 mil split-vessel.

These materials were classified into three classes regarding their FFc: cohesive, easy-flowing and free-flowing materials belong to class 1 (FFc < 4), class 2 (4 < FFc < 10), to class 3 (FFc > 10) respectively^[3]. Generally, for the materials to be suitable for DC, they should be free flowing (class 3). A Principal Component Analysis (PCA), as an unsupervised ML algorithm, was carried out on the data as a dimensional reduction analysis. The ML algorithms selected for classification problems were Random Forest (RF), Neural Network (NN), Logistic Regression (LR), Support Vector Machine (SVM), k-Nearest Neighbor (kNN), Naïve Bayes (NB), and AdaBoost (AB). Furthermore, they were evaluated by 10-fold cross-validation and with an external data set. For the external validation, 9 materials were left out of the data set used to build the model; 3 of them were cohesive (class 1), 3 were easy flowing (class 2) and the remaining 3 materials were free flowing (class 3). Only particle size and shape measurements were fed into the model to validate it.

Results

The data set used to build the ML model included 115 fully characterised materials. Regarding particle size distribution, 42 of the materials had a D₅₀ smaller than 100 µm, 56 had a D₅₀ between 100 µm and 250 µm, and 17 had a value of D₅₀ greater than 250 µm. Concerning particle shape, most of the materials’ aspect ratio falls between 0.6 and 0.8, which corresponds with a cubic shape ^[4]. Based on the FFc, the 27 materials were classified into class 1 (cohesive materials with FFc < 4), 34 were classified into class 2 (easy-flowing materials, 4 < FFc < 10), and 54 materials were classified into class 3 (free-flowing materials, FFc > 10). The physical particle properties were ranked according to their correlation with the target variable, resulting particle size distribution (D₁₀ and D₅₀) and aspect ratio distribution (a90) the variables that provided more information.

Among the models selected, RF exhibited the best performance statistics, with a value of Area under ROC of 0.807, and a value of classification accuracy of 0.67. Regarding the external validation, AB was the algorithm that could predict the class of more materials correctly (8 out of 9), followed by RF and kNN, that predicted the class of 6 materials correctly.

Implications

The ML model’s implementation enables the prediction of the material flow properties (FFc) from size and shape enabling rapid decision-making regarding manufacturing route selection. This could of course be extended to inform formulation optimization or even to provide a performance target for particle engineering efforts. The accuracy and precision of the model is reasonable; however, it can be improved with the availability of more data from a wider training set of materials. This work suggests that particle size and shape distribution are sufficient to enable prediction of flow properties. Particle size distribution, particularly D₁₀, is the variable that has the biggest impact.

References

1. Schaller, B. E., Moroney, K. M., Castro-Dominguez, B., Cronin, P., Belen-Girona, J., Ruane, P., ... & Walker, G. M. (2019). Systematic development of a high dosage formulation to enable direct compression of a poorly flowing API: A case study. International journal of pharmaceutics, 566, 615-630.
2. Trementozzi, A. N., Leung, C. Y., Osei-Yeboah, F., Irdam, E., Lin, Y., MacPhee, J. M., ... & Zawaneh, P. N. (2017). Engineered particles demonstrate improved flow properties at elevated drug loadings for direct compression manufacturing. International journal of pharmaceutics, 523(1), 133-141.
3. Zegzulka, J., Gelnar, D., Jezerska, L., Prokes, R., & Rozbroj, J. (2020). Characterization and flowability methods for metal powders. Scientific Reports, 10(1), 1-19.
4. Yu, W., Liao, L., Bharadwaj, R., & Hancock, B. C. (2017). What is the “typical” particle shape of active pharmaceutical ingredients?. Powder Technology, 313, 1-8.