(665g) Advanced Control Strategies for Continuous Pharmaceutical Manufacturing Via Regularized Dynamic Mode Decomposition with Control

The pharmaceutical industry has seen gradual adoption of continuous manufacturing (CM) for oral solid dosages (OSD) over the past 15 years, yet challenges persist due to entrenched batch manufacturing infrastructure and conservative quality systems. Traditional batch processes involve stepwise production with intermediate quality checks, but reliance on final product testing poses risks of lot-wide failures, recalls, and resource waste. In contrast, CM offers enhanced quality management through real-time monitoring and adaptive control strategies, reducing the need for scale-ups and enabling consistent product quality (Mesbah et al., 2017). However, CM requires a shift in operational paradigms, including new material tracking methods, in-line monitoring tools, and strategies to manage disturbances without interrupting production. While the transition demands upfront effort and cultural adaptation, the long-term benefits—such as improved efficiency, reduced costs, and greater flexibility—are well-documented by early adopters and supported by regulatory bodies like International Council for Harmonisation (ICH).

To make progress in this direction, quality-by-design (QbD) and quality by control (QbC) strategies as well as the appropriate use of mathematical modelling has been recommended by the ICH Q13 guideline for the development, implementation, and lifecycle management of PCM (Burcham et al, 2018). As part of QbC, active control of PCM processes is challenging due to the complex nature of process dynamics. Model predictive control (MPC) is an advanced control technique that can accommodate process constraints and complex processes involving multiple process inputs and output (Nașcu et al., 2013). As the MPC relies heavily on the predictions of the dynamic model for computation, a highly accurate model is vital for satisfactory control performance (Jelsch et al., 2021). Traditionally, these models are derived from fundamental principles such as mass and energy balance equations. However, the derivation of these first-principles models can often be tedious and costly, especially for complex nonlinear systems (Ierapetritou et al., 2017). Data-driven methods that construct dynamical models from data pose as an alternative to the theoretical approach.

Surrogate reduced order modelling provides an entirely data-driven approach to represent highly reliable yet computationally intensive models in a lower-dimensional space. These models can be used for advanced process control when meeting three essential criteria predictive accuracy, robustness and computational speed (Pantelides and Pereira, 2024). Advancements in these techniques, such as Dynamic Mode Decomposition with control (DMDc), are generating new opportunities for computationally efficient and explainable model development in comparison with complex physics-based models. DMDc is a modal decomposition technique that extracts dynamically relevant spatial structures and whose goal is to disambiguate between the underlying dynamics and the effects of actuation (Vega-Zambrano et al., 2025).

To this end, in an offline stage, we develop a regularized DMDc model, which improves upon standard DMDc (Proctor et al., 2016) by incorporating Tikhonov-regularized pseudoinversion preventing overfitting when handling ill-conditioning datasets. Data that was collected from simulations using a gPROMS FormulatedProducts® process model flowsheet of integrated twin-screw granulation (TSG) and fluidized bed drying (FBD) which was customized for the GEA ConsiGma™-25 Continuous Tableting Line CTL, at the Diamond Pilot Plant (DiPP) of the University of Sheffield (Wang et al., 2022).

Our model demonstrates low computational complexity while effectively capturing nonlinear dynamics with significant improvements observed in the performance metrics (e.g., R² > 0.93 for mean granule size prediction), when compared with state-space models obtained with N4SID tool for Multivariable Output Error State Space (MOESP) from the MATLAB system identification toolbox and Sparse Identification of Nonlinear Dynamical systems with control (SINDyc) in Python, including an enhanced uncertainty propagation performance (e.g., average length of the 95% predictive intervals = 57.85 vs. 424.66 for SINDyc and 1548.82 for MOESP).

As part of a closed-loop workflow, the developed and validated DMDc models for TSG and loading stage of the FBD were integrated into a linear model predictive control in an online stage through data exchanges between gPROMS, Python and General Algebraic Modelling System (GAMS®) using the packages gO:Python and GAMSPy. We evaluate the controller performance with setpoint tracking and disturbance rejection studies. Results indicated high accuracy in real-time monitoring and control of granule size. Additionally, we apply DMDc to the pure drying stage of the FBD using experimental data from the plant and perform open-loop MPC studies in GAMS®.

This study offers a new and interpretable advanced control strategy for CM. By integrating regularized DMDc with MPC, we provide a robust framework that aligns with QbD and QbC principles. We demonstrate the potential for real-time release testing, reduced reliance on end-product testing, and improved process control, supporting the adoption of CM in the pharmaceutical industry. This new understanding should help to improve modelling and control of PCM without hindering trust and operational agility, which has the potential to accelerate regulatory approval of the models used for control.

Acknowledgements

Financial support from EPSRC grants EP/V051008/1, EP/V034723/1, EP/W003317/1 and Peruvian Government through PRONABEC (National Program of Scholarship and Educational Loan) is gratefully acknowledged.

References

Ierapetritou, M., Sebastian Escotet‐Espinoza, M., Singh, R. (2017). Process Simulation and Control for Continuous Pharmaceutical Manufacturing of Solid Drug Products, In: Tekin, F., Schönlau, A. (Eds.), Continuous Manufacturing of Pharmaceuticals. Wiley, pp. 33–105.

Jelsch, M., Roggo, Y., Kleinebudde, P., & Krumme, M. (2021). Model predictive control in pharmaceutical continuous manufacturing: A review from a user’s perspective. Eur J Pharm Biopharm., 159, 137-142.

Mesbah A, Paulson JA, Lakerveld R, Braatz RD. (2017). Model predictive control of an integrated continuous pharmaceutical manufacturing pilot plant. Org. Process Res. Dev. 21, 844–854

Nașcu I., Diangelakis N.A., Muñoz S.G., Pistikopoulos, E.N. (2023). Advanced model predictive control strategies for evaporation processes in the pharmaceutical industries. Comput Chem Eng 173, 108212.

Pantelides, C.C., Pereira, F.E. (2024). The future of digital applications in pharmaceutical operations. Curr Opin Chem Eng.

Proctor JL, Brunton SL, Kutz JN. (2016). Dynamic mode decomposition with control. SIAM. 15, 142–161

Vega-Zambrano C, Diangelakis NA, Charitopoulos VM. (2025). Data-driven model predictive control for pharmaceutical continuous manufacturing. Int J Pharm 125322 https://doi.org/10.1016/j.ijpharm.2025.125322

Wang, L.G., Omar, C., Litster, J., Slade, D., Li, J., Salman, A., Bellinghausen, S., Barrasso, D., Mitchell, N.( 2022). Model driven design for integrated twin screw granulator and fluid bed dryer via flowsheet modelling. Int J Pharm 628, 1221