Machine learning real-time control of continuous granulation process

Abstract

The transition from traditional batch to continuous pharmaceutical manufacturing puts additional demands on the efficient process development and operation. The comprehensive understanding of complex interdependencies between critical process parameters (CPPs) and critical material attributes (CMAs) for the plants consisting of several unit operations is very challenging for process operators and experts. Therefore, the development of computational models is necessary to implement active process control and ensure a control state. Here, we present a machine-Learning (ML) based approach to build a data-driven process model and to implement real-time process control for a continuous wet granulation line. The analysis of historical process data, where a set of experiments was performed for a targeted collection of new data, has allowed us to successfully build an ML kernel and to implement a control system for a granulation plant. Furthermore, to support the training process, the process data was extended with mechanistic models implemented as soft-sensors, resulting in a hybrid model architecture. The performed tests have shown that the proposed strategy and the developed ML system can be efficiently used to perform real-time control of the continuous plant and to achieve desired CMAs such as size and loss on drying of the final granules by adjusting CPPs.

Introduction

Continuous pharmaceutical manufacturing (CPM) is considered today as a viable alternative to the traditional batch processing route to pharmaceutical manufacturing (Vanhoorne and Vervaet, 2020). CPM involves the continuous feeding, mixing, and processing of raw materials in a process train and the removal of the finished (or intermediate) product from it (Allison et al., 2015). The attractiveness of continuous processing to the pharmaceutical industry can be attributed to the improvements it brings such as: safer and faster processing due to more integrated processing units, cost saving benefits, smaller plant footprint, more opportunities to leverage process analytical technologies and real-time process monitoring (Allison et al., 2015, Schaber et al., 2011). Despite these advantages, the implementation of CPM is more complex compared to the batch operation mode and requires a paradigm shift in the control strategy (Dahlgren et al., 2019). A key element of the control strategy of continuous processes are control systems (Allison et al., 2015, Myerson et al., 2015).

In order to fulfill the high quality standards in the pharmaceutical industry, control systems must ensure that the specifications of critical material attributes (CMAs) and critical quality attributes (CQAs) are consistently achieved during the course of continuous processing (Baxendale et al., 2015, Myerson et al., 2015). By doing so, the continuous process remains in a state-of-control (Allison et al., 2015, Pantelides and Pereira, 2024), which is defined by the ICH Q10 as “a condition that provides assurance of continued process performance and product quality” (ICH, 2008). Furthermore, the ICH Q13 recommends that key elements of a control strategy for CPM involves the monitoring of the state-of-control, and when necessary, taking necessary action to ensure that the process remains in control (ICH, 2022). Therefore, the implementation of such science-based control systems to continuous pharmaceutical processes would require the integration of process analytical technology (PAT), data, control systems, and predictive models (Destro and Barolo, 2022, Ntamo et al., 2022). Due to the importance of control systems for ensuring the state-of-control of continuous pharmaceutical processes, a number of studies have been carried-out to implement such control strategies, ranging from residence time distribution (RTD)-based control (Kruisz et al., 2017, Martinetz et al., 2018), multivariate statistical process control (MSPC) (MacGregor and Kourti, 1995), hierarchical control (Stephanopoulos and Ng, 2000), classical PID control (Singh et al., 2013) to model predictive control (MPC) (Huang et al., 2021, Jelsch et al., 2021).

For instance, Pauli et al. (2020) induced the active pharmaceutical ingredient (API) content and utilized NIR spectroscopy to analyze the RTD of the material passing through an integrated continuous wet-granulation line. They showed that RTD could serve as a viable process control approach to detect the system health of the continuous granulation line even during routine production. Jelsch et al. (2022) utilized the RTD as a “process fingerprint” to monitor the system dynamics of a continuous wet granulation line. Their approach involved the use of RTD to detect changes in the process and the diversion of out-of-specification materials. They did not consider the active control of the process in the presence of process drifts, though. Bhaskar and Singh (2019) developed an RTD-based control system to divert non-conforming tablets in real-time. Their control system was developed for a continuous direct compaction tablet manufacturing process. They showed that their RTD-based control performed better than a fixed window control approach. Nevertheless, their work was mainly in silico and was not implemented in an actual pilot plant. Furthermore, RTD-based approaches are not able to capture certain material disturbances such as segregation or clogging at intermediate process points (Celikovic et al., 2024a).

Silva et al. (2017) developed an MSPC approach for a ConsiGma-25 continuous manufacturing line consisting of a twin screw high shear granulator, a fluid bed granulator and a product control unit at steady state conditions. The approach was based on a principal component analysis (PCA) model which was trained on normal operatin conditions (NOC) runs with only process variables. The results of the PCA model were depicted on Hotelling’s and control charts (Qin, 2003) to monitor the process. They successfully demonstrated that their MSPC approach could capture the process dynamics and detect all anomalies in the process. As a sequel to their work, Silva et al (2018) used batch statistical process monitoring to monitor and analyze an integrated, twin screw granulation and fluid bed granulation process. There, an orthogonal partial least squares model was used, and the focus was on pair-wise, cross-correlations between process variables with time lags. Zomer et al. (2018) demonstrated the industrial application of MSPC to monitor a ConsiGma-25 continuous tablet production line in order to detect process drifts and facilitate timely intervention. While these works show that MSPC can be used to monitor and detect faults in CPM processes, they are not able to suggest corrective actions once faults are identified.

Furthermore, MSPC approaches are typically focused on detecting faults via the process variables and do not consider CQAs. Lakerveld et al. (2013) utilized a hierarchical decomposition approach to develop a plant-wide control strategy for a model of an integrated continuous manufacturing (ICM) pilot plant. Inspired by the in silico results by Lakerveld et al., 2013, Lakerveld et al., 2015 applied the hierarchical control approach to the actual ICM plant and showed that it is able to control the critical material attributes of slurries of intermediate compounds during API synthesis steps (in crystallizers and buffer tanks). Subsequently, Mesbah et al. (2017) developed a plant-wide MPC for the ICM plant. They considered two MPC designs – subspace identification from data generated by a plant simulator, and linearization of a first principles model of the plant. They showed that both design approaches lead to comparable closed-loop performance, and the MPC is effective for controlling CQAs.

Huang et al. (2024) developed a hybrid model based on Johanson’s model (Johanson, 1965) and utilized it for the nonlinear model predictive control (NMPC) of a dry granulation process – roller compaction. Bacone et al. (2019) developed a dynamic First-Order Plus Dead Time (FOPDT) model (Bacone et al., 2020) to predict the mass flow rate of powder leaving a twin screw feeder. They also investigated the use of nonlinear model predictive control (NMPC) to maintain the powder mass flow rate in the feeder. They showed that the FOPDT model has good agreement with experimental data, and that the NMPC is able to track the set point of the powder mass flow rate even in the presence of nonlinearities and time delays.

Celikovic et al. (2023) developed a dynamic model granule size of the wet granules in a ConsiGma-25 plant by using so-called Local Linear Model Tree (LoLiMoT) algorithm (Nelles, 2001). They also developed chemometric models to monitor the API-and-liquid content of the wet granules via Raman spectroscopy. Based on these models, they developed a model predictive controller to control the API-and-liquid content (Celikovic et al., 2024a) and the particle size distribution of the wet granules processed in the twin screw granulator (Celikovic et al., 2024b). Recently, Vega-Zambrano et al. (2025) demonstrated the feasibility of dynamic mode decomposition with control (DMDc) model in designing an MPC of a twin screw granulation process. While these works are promising, they focused solely on the control of the twin screw granulator without considering the integrated system of twin screw granulation and fluid bed drying.

As the reader might have observed from the reviewed papers, studies on the application of MPC to continuous pharmaceutical manufacturing (CPM) in the literature typically involve a single unit operation (Jelsch et al., 2021, Celikovic et al., 2024b, Vega-Zambrano et al., 2025). Even in a case where MPC was applied to a CPM process consisting of integrated process units such as in Mesbah et al. (2017), the work was solely a model-based analysis and was not experimentally validated in an actual plant.

In this work, we present an industrial application of a deep learning-based MPC system to an integrated, continuous wet-granulation process for solid oral dosage forms. Our Advanced Process Control (APC) system can be classified as an Intelligent AI-based control system as coined by Huang et al. (2021). The continuous granulation plant considered in this work is a Glatt MODCOS system integrating twin screw granulation and fluid bed drying (Menth et al., 2020). Our AI system for controlling the continuous granulation plant is a deep learning model (Roggo et al., 2020) inspired by the transformer architecture (Vaswani et al., 2017). We have also infused physics-based constraints into our control system to ensure that the continuous granulation process remains in a state-of-control and recommends physically meaningful control actions (Destro et al., 2020, Huang et al., 2021).

To the best of our knowledge, we show for the first-time the successful development and application of a deep learning-based automatic control system to an integrated, continuous granulation plant (with multiple unit operations) in a pharmaceutical industrial setting (Schneider et al., 2024). Furthermore, we show that our system is able to control both particle size distribution and the moisture content of the final granules. The rest of the paper is organized as follows: in Section 2, we provide details about the continuous granulation plant and its subunits. Section 3 describes the available data was processed and enhanced with additional phenomenological aspects. Next, Section 4 presents some general discussion of data-driven MPC. In Section 5, we present the model-based control algorithm and the control system architecture. Following this, we present the results and performance of the automatic control system in Section 6. Finally, conclusions will be presented in Section 7.

This contribution focuses exclusively on the technical aspects, including system architecture, machine learning methods, integration of mechanistic model, etc. Regulatory aspects related to system validation and its potential application in GMP manufacturing are not addressed in this work.

MODCOS plant

In this contribution, a Machine Learning system for process control has been developed for a MODCOS continuous granulation line (Glatt GmbH) Germany). This plant consists of a gravimetric powder feeder, a twin screw granulator (TSG), a multi-chambered fluidized bed dryer (FBD) GPCG10, a pneumatic conveying system as well as two valves for the product discharge.

Maksym Dosta, Moritz Schneider, Christopher W. Geis, Lukas Schulte, Jan M. Kriegl, Alberto M. Gomez, Enric D. Domenech, Judith Stephan, Martin Maus, Victor N. Emenike, Machine learning real-time control of continuous granulation process, International Journal of Pharmaceutics, 2025, 126244, ISSN 0378-5173, https://doi.org/10.1016/j.ijpharm.2025.126244.