(364y) Data-Driven and Physics-Informed Machine Learning Applications in Real-Time Decision Making and Predictive Control

Efficient real-time decision-making and predictive control of process systems hinge on accurate dynamic system identifications. Traditional modeling approaches often depend on first-principles models, which require extensive domain knowledge of the process and are computationally expensive. Machine learning (ML) methods offer a robust black-box modeling alternative by capturing hidden patterns in process data without the limitations of first-principle models. For ML models to be effectively applied in real-time procedures, they must learn the temporal correlations of the process while considering other inputs. Therefore, Deep learning (DL), a subset of ML empowered by Artificial Neural Networks (ANNs), provides a reliable approach to dynamically model complex, nonlinear processes with highly relative accuracy. Advanced DL models, such as recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and self-attention transformers, have shown significant promise in the dynamic modeling of processes. These advanced models can capture both temporal dependencies within the process data and provide a reliable dynamic representation of the processes, essential for real-time predictive control and decision-making. Consequently, DL-based models are increasingly adopted across various industries for system identification, monitoring, and control, paving the way for smarter and more autonomous process systems.

Leveraging the advantages of DL methods for dynamic process modeling, we focused on developing an efficient DL-based dynamic modeling framework for integration into real-time predictive control strategies across diverse process systems. We developed a DL-based model reduction framework of an ammonia synthesis packed-bed reactor in chemical process systems for further dynamic optimization. Using offline computational fluid dynamics (CFD) simulations of the reactor’s multi-scale model, we developed a robust and fast dynamic surrogate model of the reactor based on LSTM-based seq2seq architecture. Subsequently, the surrogate model was integrated into a model predictive control (MPC) framework to control the maximum temperature of the reactor and outlet ammonia concentration based on inlet flow and removed heat from the reactor. The surrogate model demonstrated promising results as an accurate dynamic representation of the multi-scale model in closed-loop simulations, alleviating the high computational demands associated with the multi-scale model. In a consecutive work, we addressed the real-time measurement limitations of control objectives and developed a soft sensor based on feed-forward neural networks (FNNs) that utilize distributed temperature profiles of the reactor as inputs to estimate the control objectives. Further closed-loop integrations of the soft sensor alongside the surrogate model showed smooth steady-state convergence, highlighting the promising application of DL methods for dynamic system identification, complex multi-scale model reduction, and accurate state estimation in real-time predictive control.

Despite the effectiveness of data-driven methods, their accuracy and development rely solely on process data, and they are not aware of the domain knowledge of process systems, which hinders their performance in unforeseen circumstances. Addressing these limitations, particularly in agricultural processes, we introduced a Physics-Informed Machine Learning (PIML) framework for dynamic soil moisture modeling. Accurate prediction of soil moisture, considering seasonality and climatic trends, especially with the effects of global warming, plays a significant role in agricultural decision-making. Therefore, the proposed PIML framework integrates soil-water physics with ML and DL techniques, such as support vector machines (SVM), random forests (RF), and LSTMs, to accurately predict soil water content. In this hybrid approach, each PIML model is trained using the Markov soil moisture prediction model as the kernel, demonstrating soil physics and the Kansas Mesonet dataset. Two types of PIML methods, with Uncertainty Quantification (UQ) and Manipulated Cost Function (MCF), are trained and validated using Kansas Mesonet datasets to find the optimal PIML approach. Comparing the PIML results with the Markov model showed significant predictive accuracy enhancements, benefiting precision agriculture and agricultural decision-making.

Research Interests: Deep-learning based process modeling , Physics-informed Nerual Networks, Process Control.