Hybrid Modeling for Process Systems Simulation: Integrating Process Understanding and Machine Learning

Recent advances in data acquisition from the production lines in the (bio)chemical industry have resulted in improved possibilities for including this data in the modeling of such processes. The usage of data for process model development is not a new concept with classical applications being data introduced through empirical correlation models and parameter estimation in first-principles models in what can be classified as gray box models. A subgroup of gray box modeling is what is known as hybrid modeling, combining process knowledge-based modeling with data-based Machine Learning. Even though extensive work has been done within the field of gray box and hybrid modeling, questions still need to be answered regarding how to select a suitable modeling approach and how these models can effectively be applied for real-time applications in industrial settings.

This thesis explores the field of gray box and hybrid modeling using various case studies. Through these case studies, recommendations on how to select an appropriate modeling approach based on the information available and the purpose of the model. Based on both pilot and industrial scale case studies, this thesis outlines considerations for implementing hybrid models in real-time applications within an industrial setting.

A simulated CSTR case and a pilot scale bubble column case were used for investigating gray box and hybrid modeling approaches. The studies employed various techniques for predictive modeling of dynamic variables, including stochastic gray box modeling, parameter fitting in empirical correlations, hybrid semi-parametric modeling, symbolic regression, and Physics-Informed learning approaches. Stochastic gray box modeling and
symbolic regression provide transparent models, allowing developers to gain a deeper understanding of the processes involved. Hybrid semi-parametric models deliver highly accurate and adaptable models by combining the strengths of both mechanistic and data driven approaches. Finally, physics-informed learning facilitates the creation of surrogate models that operate with high computational efficiency, making them suitable for handling computationally expensive systems. The investigation resulted in a gray box model selection framework, which provides general guidelines for selecting an appropriate method for predictive modeling of dynamic variables.

Implementation of hybrid model-based real-time applications in an industrial setting was investigated by introducing a digital twin for the pilot scale bubble column and real-time optimization for an industrial scale malting process. For such an endeavor to be successful, the model should be developed as deployable software and the full lifetime of the application should be taken into account. A modular software approach was used for implementing an ensemble hybrid semi-parametric model for the pilot-scale bubble column unit, incorporating real-time batch-incremental learning. This approach allowed the model to be updated whenever new data became available, significantly enhancing its performance. In the industrial malting process case study, a systematic framework for model development and production deployment was applied to introduce energy and process real-time optimization of the germination stage. A pragmatic approach was used in developing the final model for the malting process, considering the practical constraints and limitations of the available data. The resulting models demonstrated good accuracy when evaluated on data similar to what was used for training. However, some of these models were challenging to justify in terms of process understanding. The models implemented for the germination stage performed poorly with out-of-domain data, highlighting the necessity of updating the model when a data domain shift occurs.

To successfully implement hybrid model-based applications in industrial settings, it is crucial to involve process operators, process engineers, and specialists who can contribute to the development of practical and useful applications. Additionally, plant leadership must be willing to prioritize the project and be open to operating the plant in a manner that differs from standard procedures. To achieve this, it is important to present successful use cases within industrial settings that demonstrate the benefits of utilizing hybrid modeling for process optimization.