Author ORCID Identifier
Semester
Summer
Date of Graduation
2024
Document Type
Dissertation
Degree Type
PhD
College
Statler College of Engineering and Mineral Resources
Department
Chemical and Biomedical Engineering
Committee Chair
Debangsu Bhattacharyya
Committee Co-Chair
Fernando Lima
Committee Member
Fernando Lima
Committee Member
Stephen Zitney
Committee Member
Yuhe Tian
Committee Member
Miguel Zamarripa
Abstract
First-principles models can provide very good predictions even for cases when there are no data at all, or data are limited in certain range of operating conditions, or for cases where data collection is infeasible. However, the development of accurate first-principles models for complex nonlinear dynamic systems can be time consuming, computationally expensive, and may be infeasible for certain systems due to lack of sufficient knowledge (information). It is also challenging to adapt first-principles models for time-varying systems. Furthermore, it can be difficult, if not impossible, to develop accurate models for some complex phenomena that are poorly understood. On the contrary, black-box or data-driven models are relatively easier to develop, simulate and adapt online. Over the last few decades, numerous advantages of using data science and artificial intelligence (AI) / machine learning (ML) approaches in the field of chemical engineering have been exploited and analyzed. Neural networks (NNs) have been identified as one of the many different artificial intelligence approaches that exhibit a strong potential to modeling highly complex nonlinear dynamic systems.
This work is primarily focused on the development, training and applications of novel deterministic and stochastic dynamic physics-constrained hybrid neural network structures and hybrid first principles-NN models for data-driven modeling of complex nonlinear dynamic chemical process systems. The NN models developed and proposed as part of this research can be either implemented independently in a fully data-driven approach or be synergistically hybridized with other first-principles (physics-based) models for exploiting their key strengths. Most of the AI/ML techniques, especially neural networks, have evolved since the last few decades for a diverse range of chemical engineering applications such as data classification, fault detection and diagnosis, chemical process design, monitoring, and control. Although a plethora of different types of deep and shallow NN models have been successfully implemented for modeling various complex nonlinear transient chemical processes, the successful development of such techniques require large data sets which may not often be available for modeling various chemical engineering systems. This work develops novel architectures and training algorithms for hybrid all-nonlinear series and parallel static-dynamic NN models for applications to chemical systems. The sequential decomposition-based training algorithms thus formulated exploit the model structure, leading to independent and separate training of static and dynamic networks by different parameter estimation algorithms while still achieving optimal convergence by solving an outer layer optimization. The all-nonlinear series and parallel networks have been shown to significantly outperform typical deep recurrent neural networks, with the proposed sequential training algorithms leading to 50-100 times faster computation than simultaneous training algorithms.
One of the more recent advancements in the field of process modeling using NNs points to the augmentation of various physics conservation laws pertaining to a system while constructing optimal network models during both training and simulation. The primary motivation of such approaches stems from the fact that the measurement / experimental data available for training may not necessarily satisfy mass and energy balance equations and/or other thermodynamics / physics based constraints. If such conservation laws are not considered during machine learning, model predictions can violate system physics and hence are not meaningful. However, in almost all existing literature on physics-informed neural networks (PINNs), the physics conservation equations have been augmented in the loss function as additional penalty terms, thus serving as soft constraints without ensuring an ‘exact’ satisfaction of the conservation laws. This work focuses on the development of algorithms for exactly satisfying physics conservation laws such as mass and energy balances as well as thermodynamics constraints during both training and forward problems using noisy transient data. The mass-energy-thermodynamics constrained neural network models developed in this research are found to be very accurate capturing the system truth with minimum bias for all examples that are evaluated, even when the training data are corrupted with uncertainties.
Another alternative approach that is commonly practiced for including mechanistic (first-principles) information along with data-driven models considers the synergistic integration of first-principles (FP) and AI models leading to the construction of hybrid first-principles artificial intelligence (FP+AI) models. This work also discusses novel approaches developed for coupling FP and AI models in series, parallel, and integrated configurations, along with algorithmic capabilities to ensure that the resulting optimal models do not violate system physics even after hybridization. Additionally, the proposed hybrid first-principles machine learning approaches, when applied to commercial power plant data, demonstrate >15% improved predictive accuracy compared to the corresponding standalone FP models. For modeling uncertain systems, probabilistic neural networks can be highly useful. However, synthesis of optimal hybrid networks and parameter estimation for probabilistic neural network models of dynamic uncertain systems without suffering from the ‘curse of dimensionality’ is considerably challenging. It is also challenging to satisfy physics constraints when probabilistic NNs are used. The final set of model architectures and algorithms developed in this research are aimed at optimal integration of stochastic and conventional NNs as well as efficient formulation of training algorithms for hybrid series and parallel constrained stochastic-deterministic network models. The corresponding approaches proposed in this work for construction of optimal parsimonious hybrid stochastic networks exhibit considerably superior results compared to the state-of-the-art approaches when evaluated for large-scale complex nonlinear process systems. All proposed data-driven / hybrid models and algorithms are evaluated and validated for modeling various complex nonlinear transient noisy chemical process systems.
Recommended Citation
Mukherjee, Angan, "Development of Dynamic Mass-Energy-Thermodynamics Constrained Hybrid Neural Network Models for Process Systems Applications" (2024). Graduate Theses, Dissertations, and Problem Reports. 12562.
https://researchrepository.wvu.edu/etd/12562
Embargo Reason
Publication Pending
Included in
Other Operations Research, Systems Engineering and Industrial Engineering Commons, Process Control and Systems Commons, Systems Engineering Commons