(567d) Hybrid Machine Learning Assisted Model Predictive Control of a Continuous Dry Granulation Tableting Line

Active control strategies are playing an essential role in modern pharmaceutical manufacturing. Instead of controlling the process input variables within tight ranges as practiced in traditional control approaches, a more flexible active control strategy has the potential to deal with process uncertainties and variations in raw material properties. Model predictive control (MPC) often serves as the preferred candidate for a centralized plant-wide control strategy. When there are many constraints on critical quality attributes and strong process variable interactions exist in the multiple-input multiple-output (MIMO) system, MPC can demonstrate better capabilities in setpoint tracking and disturbance rejection compared to classical PID control. However, the implementation of MPC in pharmaceutical tablet manufacturing processes is still in infancy. Most work reported in the literature is limited to a single process unit or the direct compaction process; integrated dry granulation or wet granulation tableting processes still require further research [1].

Dry granulation via roller compaction (RC) is increasingly applied in the pharmaceutical industry to change particle size distribution (PSD) of in-process powders or final products. The granulation step is beneficial for improving powder flowability and addresses problems related to dustiness, low bulk density, and poor blend uniformity [2]. Nevertheless, unless the process is well-controlled, when granules are subsequently compressed into tablets, the powder compressibility may be compromised, and tablet tensile strength is sacrificed [3]. Therefore, a model to correlate the properties of powders, ribbons, granules, to those of the tablets produced is required to understand and further control the process appropriately. Mechanistic models such as Johanson’s model [4] and Reynold’s model [5] are typically used to describe roll compaction, while population balance models (PBM) can account for the milling step. However, it is complicated to determine the breakage function in the PBM purely based on ribbon fracture physics. Machine learning (ML) is a preferred data-driven alternative to developing a mechanistic model. Moreover, ML and mechanistic model components can be combined into a hybrid model to maintain high physical interpretability and feasibility.

This study is focused on developing and implementing model predictive control in continuous dry granulation tableting processes. First, a hybrid ML-assisted model is reported, which serves to predict the process behavior and the essential product attributes. Process behavior of the rotary tablet press and tablet properties are described by a mechanistic model which has been implemented in previous work [6]. Neural network and genetic programming approaches are used to predict the PSD of granules produced by the roller compactor, as well as to map the impact of granule PSD and roll compaction on tablet properties. In the second part, the hybrid ML-assisted MPC, implemented both in silico and in our physical continuous tableting pilot plant, are described. The results of investigating alternative control strategies and their real-time control performance will be summarized.

[1] Jelsch, M., Roggo, Y., Kleinebudde, P., & Krumme, M. (2021). Model predictive control in pharmaceutical continuous manufacturing: A review from a user’s perspective. European Journal of Pharmaceutics and Biopharmaceutics, 159, 137-142.

[2] Toson, P., Lopes, D. G., Paus, R., Kumar, A., Geens, J., Stibale, S., ... & Khinast, J. (2019). Model-based approach to the design of pharmaceutical roller-compaction processes. International journal of pharmaceutics: X, 1, 100005.

[3] Herting, M. G., & Kleinebudde, P. (2008). Studies on the reduction of tensile strength of tablets after roll compaction/dry granulation. European journal of pharmaceutics and biopharmaceutics, 70(1), 372-379.

[4] Johanson, J. R. (1965). A rolling theory for granular solids.

[5] Reynolds, G., Ingale, R., Roberts, R., Kothari, S., & Gururajan, B. (2010). Practical application of roller compaction process modeling. Computers & chemical engineering, 34(7), 1049-1057.

[6] Huang, Y. S., Sheriff, M. Z., Bachawala, S., Gonzalez, M., Nagy, Z. K., & Reklaitis, G. V. (2021). Evaluation of a Combined MHE-NMPC Approach to Handle Plant-Model Mismatch in a Rotary Tablet Press. Processes, 9(9), 1612.