| dc.description.abstract | Recently, there has been growing interest in data-based modeling as the amount of data available has increased tremendously. One such method is Dynamic Mode Decomposition with Control technique, which builds temporally local linear models using data. But its limited domain of applicability (DA) hinders its use for prediction purposes. To overcome this challenge, we proposed an algorithm that utilizes multiple "local" training datasets, and it was applied successfully to hydraulic fracturing. Although data-based modeling offers simplicity and ease of construction, it lacks robustness and parametric interpretability, unlike first-principles modeling.
To balance the advantages and disadvantages of data-based models and first-principles models,
hybrid modeling was proposed using artificial neural networks (ANNs). Since then, Machine
Learning (ML) has advanced where deep neural networks (DNNs) with more than three layers can be trained to approximate any function accurately. In this work, we proposed a deep hybrid modeling (DHM) framework that integrates first-principles with DNNs and successfully applied it to two complex processes, i.e., hydraulic fracturing and full-scale fermentation reactor. Similarly, Universal Differential Equations (UDEs) was proposed in ML where DNNs are represented as ODEs and solved using ODE solvers. We utilized UDEs to successfully build a DHM using simulation and experimental data for batch production of ϐ-carotene. One limitation of DHM is that its DA is affected by the DNN within it, and its accuracy is high within its DA. Therefore, it is important to consider its DA when designing a model-based controller. To this end, we proposed a Control Lyapunov-Barrier Function (CLBF)-MPC to stabilize and ensure that the closed-loop system stays within DA of DHM. Theoretical guarantees were provided for the CLBF-MPC controller, and it was successfully implemented on a CSTR. The idea of integrating physics with ML can be extended to Reinforcement Learning (RL). In case when model-based controller design is not possible, we proposed a model-free Deep RL (DRL) controller that utilizes prior knowledge in its reward function to quicken the learning process. This DRL controller was successfully applied to hydraulic fracturing wherein Nolte’s law was included in the reward function for fast convergence. | |
