(287c) Implementation of Hybrid Models to Perform System Analyses with Model Maintenance in Continuous Pharmaceutical Manufacturing

Recent development of Industry 4.0 technologies, including the advancements of cyber-physical systems (CPS) and Internet of Things (IoT), have increased the amount of data available from manufacturing shop floors and analytical laboratories^1-5. The abundance of data contributes to the emergence and improvement of data analytic methods and cloud computing platforms, aiming to extract more information and knowledge from data to support predictive modeling and decision making^{6, 7}. Traditionally, process industries, including continuous pharmaceutical manufacturing (CPM), mainly use white-box (WB) models for predictive modeling⁸. These WB models, including mechanistic, phenomenological, and first-principle models, rely heavily or solely on physical understandings of the system. Though the WB models provide a high degree of transparency on addressing the physical significance of variables and model structures, the development of the models is costly, and the prediction performance is largely dependent on the mastery of full process knowledge. To address these challenges utilizing the vast availability of data, recent efforts have focused on combining data with process knowledge to form hybrid models (HMs). HMs consist of at least one white-box and one black-box (BB) sub-model, where the BB sub-models take in historical process data and adopt different machine learning (ML) algorithms like artificial neural network (ANN) and support vector machine (SVM)⁹. The development of HMs of various structures provides greater flexibility and better predictability in complex engineering systems, making it applicable in modeling CPM processes.

Current research has been dedicated to developing proof-in-concept HMs for individual unit operations of CPM processes, but the capability of HMs to perform system analyses like sensitivity and feasibility analysis remains unexplored. These analyses can identify critical process parameters and design space of CPM systems, which is crucial for process design and operations¹⁰. Another challenge of implementing HMs in CPM systems is related to model maintenance. Parts of HMs are data-driven BB sub-models, and the offline training of BB sub-models with defined sets of historical data leads to static BB models¹¹. These static BB sub-models only reflect the system at the timepoint that the models are developed, and they will not change unless the models are re-trained with new data and algorithms¹². As time progresses, assuming the process stays the same, new datasets collected from plants should have the same or similar underlying distribution as the data used to train the original model. Using these new datasets to re-train should yield models with comparable performance to their predecessors. However, real datasets from manufacturing plants often contain normal operational data and abnormal data (e.g. highly noisy data, faulty data, missing data points, testing and maintenance data), and the quality of these new data can affect the re-training result considerably. On the other hand, like in the majority of manufacturing processes, a shift in the distributions of manufacturing data from CPM plants can happen over time due to changes such as environmental conditions, equipment performance, and control strategy. Because these changes are not captured in the model training process, deteriorating performance of HMs can be observed, which is a classic scenario known as model drift¹¹. Model maintenance strategies therefore need to be in place to capture this shift and to ensure the accuracy of model predictions.

In this work, algorithms are developed to perform system analyses using HMs and to address model maintenance issues in CPM. For system analyses, HMs developed for individual unit operations of the continuous direct compaction (DC) line are connected in a flowsheet model, and they are then used to perform sensitivity and feasibility analysis^{10, 13, 14}. The results from HMs are compared to those produced by WB models to compare the model performance. Next, two tasks are carried out to explore model maintenance schemes for HMs. Different types and amounts of new datasets are first fed into the re-training process to study the impact of the quality of additional data to HM performance. This analysis can yield practical guidance in data handling and maintenance frequency based on characteristics of newly available datasets. Then, to address the issue of model drift, different learning techniques and adaptive methods for model maintenance are developed. Manual re-training, incremental learning, and continuous streaming procedures are explored along with blind and informed adaptive algorithms^{11, 15, 16}. Blind adaptations update an HM continuously without realizing that a drift exists, whereas informed algorithms only become effective when predefined triggers are activated¹¹. The results will provide insights on how model maintenance for HMs can be performed. Together with system analysis capabilities, this will allow for broader acceptance of HMs for practical use in CPM with the development of Industry 4.0 framework.

1. Ding, B., Pharma Industry 4.0: Literature review and research opportunities in sustainable pharmaceutical supply chains. Process Safety and Environmental Protection 2018, 119, 115-130.
2. Cao, H.; Mushnoori, S.; Higgins, B.; Kollipara, C.; Fermier, A.; Hausner, D.; Jha, S.; Singh, R.; Ierapetritou, M.; Ramachandran, R., A Systematic Framework for Data Management and Integration in a Continuous Pharmaceutical Manufacturing Processing Line. Processes 2018, 6 (5).
3. Reis, M. S.; Kenett, R., Assessing the value of information of data-centric activities in the chemical processing industry 4.0. AIChE Journal 2018, 64 (11), 3868-3881.
4. Barenji, R. V.; Akdag, Y.; Yet, B.; Oner, L., Cyber-physical-based PAT (CPbPAT) framework for Pharma 4.0. Int J Pharm 2019, 567, 118445.
5. Steinwandter, V.; Borchert, D.; Herwig, C., Data science tools and applications on the way to Pharma 4.0. Drug Discov Today 2019, 24 (9), 1795-1805.
6. Thomas, M. C.; Zhu, W.; Romagnoli, J. A., Data mining and clustering in chemical process databases for monitoring and knowledge discovery. Journal of Process Control 2018, 67, 160-175.
7. Liu, Y.; Xu, X., Industry 4.0 and Cloud Manufacturing: A Comparative Analysis. Journal of Manufacturing Science and Engineering 2016, 139 (3).
8. Zendehboudi, S.; Rezaei, N.; Lohi, A., Applications of hybrid models in chemical, petroleum, and energy systems: A systematic review. Applied Energy 2018, 228, 2539-2566.
9. von Stosch, M.; Oliveira, R.; Peres, J.; Feyo de Azevedo, S., Hybrid semi-parametric modeling in process systems engineering: Past, present and future. Comput Chem Eng 2014, 60, 86-101.
10. Wang, Z.; Escotet-Espinoza, M. S.; Ierapetritou, M., Process analysis and optimization of continuous pharmaceutical manufacturing using flowsheet models. Comput Chem Eng 2017, 107, 77-91.
11. Gama, J.; Žliobaitė, I.; Bifet, A.; Pechenizkiy, M.; Bouchachia, A., A survey on concept drift adaptation. ACM Computing Surveys 2014, 46 (4), 1-37.
12. Webb, G. I.; Hyde, R.; Cao, H.; Nguyen, H. L.; Petitjean, F., Characterizing concept drift. Data Mining and Knowledge Discovery 2016, 30 (4), 964-994.
13. Metta, N.; Ghijs, M.; Schäfer, E.; Kumar, A.; Cappuyns, P.; Assche, I. V.; Singh, R.; Ramachandran, R.; Beer, T. D.; Ierapetritou, M.; Nopens, I., Dynamic Flowsheet Model Development and Sensitivity Analysis of a Continuous Pharmaceutical Tablet Manufacturing Process Using the Wet Granulation Route. Processes 2019, 7 (4).
14. Boukouvala, F.; Niotis, V.; Ramachandran, R.; Muzzio, F. J.; Ierapetritou, M. G., An integrated approach for dynamic flowsheet modeling and sensitivity analysis of a continuous tablet manufacturing process. Comput Chem Eng 2012, 42, 30-47.
15. Kadlec, P.; Grbić, R.; Gabrys, B., Review of adaptation mechanisms for data-driven soft sensors. Comput Chem Eng 2011, 35 (1), 1-24.
16. Pechenizkiy, M.; Bakker, J.; Žliobaitė, I.; Ivannikov, A.; Kärkkäinen, T., Online mass flow prediction in CFB boilers with explicit detection of sudden concept drift. ACM SIGKDD Explorations Newsletter 2010, 11 (2), 109-116.