(335a) Integration of High-Fidelity Simulation and Surrogate Modeling for Reduced Model Development in Continuous Pharmaceutical Unit Operations

Developing accurate prediction of continuous manufacturing (CM) lines has been an active research area for pharmaceutical industry with economic incentives and government initiatives to manufacture high quality products [1, 2]. Process models used to obtain the prediction for unit operations are commonly developed using system-level process knowledge and experimental data, which provide overall systemic information. However, this can limit the insight into the unit dynamics and hinder the prediction of important critical quality attributes [3, 4]. In CM routes for solid-based products, continuous powder blender plays a central role, providing mixing of ingredient powders. For blenders, the important quality attributes (CQA) include blend uniformity (BU) and residence time distribution (RTD) [5]. Powder holdup within the unit is another important attribute that impacts material flow in the blender, along with BU and RTD. To model this unit operation, a popularly used approach involves use of RTD models [6, 7] which lacks calculation of BU and holdup variability. Recent studies have explored the prediction of BU using compartment modeling strategies [8], but these models do not capture the variation of powder holdup along the blender [9] along its effects on CQAs [10]. To address these drawbacks and develop accurate models for unit operations, the proposed work aims at developing a reduced model based on integration of high-fidelity simulations, that can incorporate detailed unit dynamics, with surrogate models to reduce the computational complexity. This approach is demonstrated for a continuous powder blender.

The proposed work is divided into two parts. The first part focuses on developing a reduced model of the powder blender using high-fidelity discrete element simulations (DEM) [11]. Initially, the entire blender unit is simulated using DEM. Based on the observed flow of powder, the unit is broken down into the inlet section, multiple intermediate sections, and the outlet section. The dynamic information of each section, such as powder holdup, blend uniformity, and residence time, is extracted. This information is subsequently used to construct a reduced process model of the blender, incorporating powder flow and mixing along the different sections of the unit. The resulting process model thus incorporates axial mixing, radial mixing, and variable holdup profiles along the length of the blender. The second part of the work aims at improving the developed process model to incorporate the effects of varying process conditions on the degree of mixing and powder flow using a surrogate model. Physics-informed neural networks (PINN) is implemented as the surrogate model, given the flexibility of this model to incorporate physics within the construction of neural network [12]. PINN follows the general structure of an artificial neural network (ANN) [13], with a modified loss function to apply penalty to the predictions that violate physical constraints such as material balances and limits of holdup variations [12]. Therefore, the predictions that violate the constraints result in an increase in the loss function, allowing the training algorithm to favor the predictions that satisfy the physical constraints. The PINN model is integrated within the developed reduced model leading to accurate predictions. The developed process model for continuous blender can thus provide accurate predictions of with low computational cost, allowing the integration into process flowsheets and the Pharma 4.0 framework. This approach can also be generalized to other processes to better utilize physical knowledge, reduce computational complexity in process simulation, and improve predictability.

References:

1. Yu, L.X. and M. Kopcha, The future of pharmaceutical quality and the path to get there. International Journal of Pharmaceutics, 2017. 528(1-2): p. 354-359.

2. O'Connor, T.F., Yu, L. X. & Lee, S. L. , Emerging technology: A key enabler for modernizing pharmaceutical manufacturing and advancing product quality. . International Journal of Pharmaceutics, 2016. 509: p. 492-498.

3. Jourdan, N., et al., Compartmental Modelling in chemical engineering: A critical review. Chemical Engineering Science, 2019. 210.

4. Rogers, A. and M. Ierapetritou, Challenges and opportunities in modeling pharmaceutical manufacturing processes. Computers & Chemical Engineering, 2015. 81: p. 32-39.

5. Engisch, W. and F. Muzzio, Using Residence Time Distributions (RTDs) to Address the Traceability of Raw Materials in Continuous Pharmaceutical Manufacturing. Journal of Pharmaceutical Innovation, 2016. 11(1): p. 64-81.

6. Escotet-Espinoza, M.S., et al., Effect of tracer material properties on the residence time distribution (RTD) of continuous powder blending operations. Part I of II: Experimental evaluation. Powder Technology, 2019. 342: p. 744-763.

7. Escotet-Espinoza, M.S., et al., Effect of material properties on the residence time distribution (RTD) characterization of powder blending unit operations. Part II of II: Application of models. Powder Technology, 2019. 344: p. 525-544.

8. Bhalode, P. and M. Ierapetritou, Hybrid multi-zonal compartment modeling for continuous powder blending processes. International Journal of Pharmaceutics, 2021. 602: p. 120643.

9. Vanarase, A.U., Design, modeling and real-time monitoring of continuous powder mixing processes. 2011.

10. Palmer, J., et al., Mapping key process parameters to the performance of a continuous dry powder blender in a continuous direct compression system. Powder Technology, 2020. 362: p. 659-670.

11. Ketterhagen, W.R., Ende, M. T. A. & Hancock, B. C. , Process modeling in the pharmaceutical industry using the discrete element method. Journal of Pharmaceutical Sciences 2009. 98: p. 442–470.

12. Sampat, C., and Rohit Ramachandran, Identification of Granule Growth Regimes in High Shear Wet Granulation Processes Using a Physics-Constrained Neural Network. Processes 2021. 9(5): p. 737.

13. Raissi, M., P. Perdikaris, and G.E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 2019. 378: p. 686-707.