Physics-informed machine learning-based modeling and control of dynamic process systems

Many of the processes in chemical engineering applications are of dynamic nature. Mechanistic modeling of these processes is challenging due to the nonlinearity and complexity including the fact that the models often comprise uncertainties. On the other hand, recurrent neural networks that can process sequential data are useful to be utilized to model dynamic processes by using the available data. Although these networks can capture the complexities, they might contribute to overfitting and require high-quality and adequate data. Physics-informed neural networks might offer promising results by overcoming the limitations associated with inadequate training. In this thesis, two different physics-informed training approaches are investigated. The first approach is using a multi-objective loss function in the training including the discretized form of the differential equation. The second approach is using a hybrid recurrent neural network cell with embedded physics-informed and data-driven nodes performing Euler discretization. Two synthetic case studies for a semi-batch reactor and a wastewater treatment plant are developed to observe the effect of physics-informed approaches. Additionally, data from a wastewater treatment plant in Tüpraş İzmit Refinery is used to observe the difficulties and usefulness of the implemented approach in industrial case studies. It is demonstrated that physics-informed neural networks can improve test performance even though a decrease in training performance might be observed. Additionally, hybrid recurrent neural networks predict the trend successfully regardless of the learning performances of the data-driven nodes. For some of the machine learning models, hyperparameter optimization is done using search algorithms. When physics-informed training is performed, smaller and more robust architectures are obtained using hyperparameter optimization. This thesis also includes the comparison of nonlinear model predictive controllers which use the first-principles model, recurrent neural network model, and physics-informed recurrent neural network model. The controller performances are observed for two case studies of a semi-batch reactor and a Van de Vusse reactor. It is concluded that recurrent neural network-based controllers could achieve desired performances and deliver similar results with nonlinear first principles-based model predictive controllers. Finally, it is observed that physics-informed recurrent neural network-based controllers could improve the controller performances compared with the physics-uninformed recurrent neural network-based controllers.