In this work, we develop a machine-learning-based predictive control design for nonlinearparabolic partial differential equation (PDE) systems using process state measurement time-series data. First, the Karhunen-Loeve expansion is used to derive dominant spatial empirical ` eigenfunctions of the nonlinear parabolic PDE system from the data. Then, these empirical eigenfunctions are used as basis functions within a Galerkin’s model reduction framework to derive the temporal evolution of a small number of temporal modes capturing the dominant dynamics of the PDE system. Subsequently, feedforward neural networks (FNN) are used to approximate the reduced-order dominant dynamics of the parabolic PDE system from the data within a desired operating region. Lyapunov-based model predictive control (MPC) scheme using FNN models is developed to stabilize the nonlinear parabolic PDE system. Finally, a diffusion-reaction process example is used to demonstrate the effectiveness of the proposed machine-learning-based predictive control method.

This work develops a computational fluid dynamics (CFD) framework for high-fidelity modeling

of steam methane reforming reactors and furnaces. Initially, a CFD model for a steam

methane reforming reactor is developed and its results are shown to closely match industrial

plant data. Subsequently, CFD models of steam methane reforming furnaces are developed

for both pilot-scale and full industrial scale furnaces. These furnace CFD models capture the

physical dimension, transport phenomena, and core components of a reformer utilized in an

industrial plant. Comparison of the CFD models with industrial plant data demonstrates that

model predictions are within 1% of industrial measurements for consistent reformer conditions.

Finally, to automate the use of the CFD models for reformer furnace balancing, we

develop a workflow for reformer simulation on the Smart Manufacturing platform. The workflow

is designed to be executed without the need of an expert user, to be deployed in a cloud

environment and to be fully or partially used.

Facilitated by the increasing importance and demand of semiconductors for the smartphoneand even the automobile industry, thermal atomic layer deposition (ALD) has gained tremendous industrial interest as it offers a way to efficiently deposit thin-films with ultra-high conformity. It is chosen largely due to its superior ability to deliver ultra-conformal dielectric thin-films with high aspect-ratio surface structures, which are encountered more and more often in the novel design of metal-oxide-semiconductor field-effect transistors (MOSFETs) in the NAND (Not-And)-type flash memory devices. Based on the traditional thermal ALD method, the plasma enhanced atomic layer deposition (PEALD) allows for lower operating temperature and speeds up the deposition process with the involvement of plasma species. Despite the popularity of these two methods, the development of their operation policies remains a complicated and expensive task, which motivates the construction of an accurate and comprehensive simulation model.

A series of studies have been carried out to elucidate the mechanisms and the conceptof the PEALD process. In particular, process characterization focuses on the development of a first-principles-based three-dimensional, multiscale computational fluid dynamics (CFD) model, together with reactor geometry optimizations, of SiO2 thinfilm thermal atomic layer deposition (ALD) using bis(tertiary-butylamino)silane (BTBAS) and ozone as precursors. Also, a comprehensive multiscale computational fluid dynamics (CFD) model incorporating the plasma generation chamber is used in the deposition of HfO2 thin-films utilizing tetrakis(dimethylamido) hafnium (TDMAHf) and O2 plasma as precursors.

Despite the great deal of research effort, ALD and PEALD processes have not been fullycharacterized from the view point of process control. This study aims to use previously developed multiscale CFD simulation model to design and evaluate an optimized control scheme to deal with industrially-relevant disturbances. Specifically, an integrated control scheme using a proportional-integral (PI) controller and a run-to-run (R2R) controller is proposed and evaluated to ensure the deposition of high-quality conformal thin-films. The ALD and PEALD processes under typical disturbances are simulated using the multiscale CFD model, and the integrated controllers are applied in the process domain. Using the controller parameters determined from the open-loop results, the developed integrated PI-R2R controller successfully mitigates the disturbances in the reactor with the combined effort of both controllers.

Big data is considered to play an important role in the fourth industrial revolution, which requires engineers and computers to fully utilize data to make smart decisions and improve the performance of industrial processes and of their control and safety systems. Traditionally, industrial process control systems rely on a (usually linear) data-driven model with parameters that are identified from industrial/simulation data, and in certain cases, for example, in profit-critical control loops, on first-principles models (with data-determined model parameters) that describe the underlying physico-chemical phenomena. However, modeling large-scale, complex nonlinear processes continues to be a major challenge in process systems engineering. Modeling is particularly important now and into the future, as process models are key elements of advanced model-based control systems, e.g., model predictive control (MPC) and economic MPC (EMPC).

Due to the wide variety of applications, machine learning models have great potential, yet, the development of rigorous and systematic methods for incorporating machine learning techniques in nonlinear process control and operational safety is in its infancy. Traditionally, operational safety of chemical processes has been addressed through process design considerations and through a hierarchical, independent design of control and safety systems. However, the consistent accidents throughout chemical process plant history (including several high profile disasters in the last decade) have motivated researchers to design control systems that explicitly account for process operational safety considerations. In particular, a new design of control systems such as model predictive controllers (MPC) that incorporate safety considerations and can be coordinated with safety systems has the potential to significantly improve process operational safety and avoid unnecessary triggering of alarms systems, where machine learning techniques can be utilized to derive dynamic process models. However, the rigorous design of safety-based control systems poses new challenges that cannot be addressed with traditional process control methods, including, for example, proving simultaneous closed-loop stability and safety. On the other hand, cybersecurity has become increasingly important in chemical process industries in recent years as cyber-attacks that have grown in sophistication and frequency have become another leading cause of process safety incidents. While the traditional methods of handling cyber-attacks in control systems still rely partly on human analysis and mainly fall into the area of fault diagnosis, the intelligence of cyber-attacks and their accessibility to control system information has recently motivated researchers to develop cyber-attack detection and resilient operation control strategies to address directly cybersecurity concerns.

Motivated by the above considerations, this dissertation presents the use of machine learning techniques in model predictive control, operational safety and cybersecurity for chemical processes described by nonlinear dynamic models. The motivation and organization of this dissertation are first presented. Then, the use of machine learning techniques to develop data-driven nonlinear dynamic process models to be used in model predictive controllers is presented, followed by the discussion of real-time implementation with online learning of machine leaning models and of physics-based machine learning modeling methods. Subsequently, the MPC and economic MPC schemes that use control Lyapunov-barrier functions (CLBF) are presented in detail with rigorous analysis provided on their closed-loop stability, operational safety and recursive feasibility properties. Next, the development of machine-learning-based CLBF-MPC schemes is presented with process stability and safety analysis. Finally, the development of an integrated detection and control system for process cybersecurity is developed, in which several types of intelligent cyber-attacks, machine learning detection methods and resilient control strategies are presented. Throughout the dissertation, the control methods are applied to numerical simulations of nonlinear chemical process examples to demonstrate their effectiveness and performance.

Atomic layer deposition (ALD) and plasma enhanced atomic layer deposition (PEALD) are the most widely utilized deposition techniques in the semiconductor industry due to their superior ability to produce highly conformal films and to deposit materials into high aspect-ratio geometric structures. Additionally, plasma enhanced ALD is able to further speed up the deposition process and to reduce the temperature requirement through the utilization of high energy particles. However, ALD and PEALD experiments remain expensive and time-consuming, and the existing first-principles based models have not yet been able to provide solutions to key process outputs that are computationally efficient, which is necessary for on-line optimization and real-time control.

Motivated by the above considerations, this dissertation focuses on addressing these issues for both ALD and PEALD. First, for ALD, the development of key components of a comprehensive simulation framework is presented. The simulation framework integrates first-principles-based microscopic modeling, input/output modeling and optimal operation of thermal atomic layer deposition (ALD) of SiO2 thin-films using bis(tertiary-butylamino)silane (BTBAS) and ozone as precursors. Specifically, we initially utilize Density Functional Theory (DFT)-based calculations for the computation of the key thermodynamic and kinetic parameters, which are then used in the microscopic modeling of the ALD process. Subsequently, a detailed microscopic model is constructed, accounting for the microscopic lattice structure and atomic interactions, as well as multiple microscopic film growth processes including physisorption, abstraction and competing chemical reaction pathways. Kinetic Monte-Carlo (kMC) algorithms are utilized to obtain computationally efficient microscopic model solutions while preserving model fidelity. The obtained kMC simulation results are used to train Artificial Neural Network (ANN)-based data-driven models that capture the relationship between operating process parameters and time to ALD cycle completion. Specifically, a two-hidden-layer feed-forward ANN is constructed to find a feasible range of ALD operating conditions accounting for industrial considerations, and a Bayesian Regularized ANN is constructed to implement the cycle-to-cycle optimization of ALD cycle time. Extensive simulation results demonstrate the effectiveness of the proposed approaches. The kMC models successfully achieves a growth per cycle (GPC) of 1.8 A per cycle, which is in the range of reported experimental values. The ANN models accurately predict deposition time to steady-state from the given operating condition input, and the cycle time optimization using BRANN model reduces the conventional BTBAS cycle time by 60%. After developing an efficient simulation framework, a more detailed study on the optimal operation policy is performed using a multiscale data-driven model. The multiscale data-driven model captures the macroscopic process domain dynamics with a linear parameter varying model, and characterizes the microscopic domain film growth dynamics with a feed-forward artificial neural network (ANN) model. The multiscale data-driven model predicts the transient deposition rate from the following four key process operating parameters that can be manipulated, measured or estimated by process engineers: precursor feed flow rate, operating pressure, surface heating, and transient film coverage. Our results demonstrate that the multiscale data-driven model can efficiently characterize the transient input-output relationship for the SiO2 thermal ALD process using Bis(tertiary-butylamino)silane (BTBAS) as the Si precursor. The multiscale data-driven model successfully reduces the computational time from 0.6 - 1.2 hr for each time step, which is required for the first-principles based multiscale computational fluid dynamics (CFD) model, to less than 0.1 s, making its real-time usage feasible. The developed data-driven modeling methodology can be further generalized and used for other thermal ALD or similar deposition systems, which will greatly enhance the feasibility of industrial manufacturing processes.

For PEALD, a similar methodology is adopted. First, an accurate, yet efficient kinetic Monte Carlo (kMC) model and an associated machine learning (ML) analysis are proposed to capture the surface deposition mechanism of the HfO2 thin-film PEALD using Tetrakis-dimethylamino-Hafnium (TDMAHf) and oxygen plasma. Density Functional Theory (DFT) calculations are performed to obtain the key kinetic parameters and the structural details. After the model is validated by experimental data, a database is generated to explore a variety of precursor partial pressure and substrate temperature combinations using the kMC algorithm. A feed-forward Bayesian regularized artificial neural network (BRANN) is then constructed to characterize the input-output relationship and to investigate the optimal operating condition. Next, based on an associated work on a comprehensive 3D multiscale computational fluid dynamics (CFD) model for the PEALD process, a 2D axisymmetric reduction of the previous 3D model of PEALD reactors with and without the showerhead design has been adopted to enhance the computational efficiency. Using the derived 2D CFD model, a data-driven model is constructed based on a recurrent neural network (RNN) for process characterization. The developed integrated data-driven model is demonstrated to accurately characterize the key aspects of the deposition process as well as the gas-phase transport profile while maintaining computational efficiency. The derived data-driven model is further validated with the results from a full 3D multiscale CFD model to evaluate model discrepancy. Using the data-driven model, an operational strategy database is generated, from which the optimal operating conditions can be determined for the deposition of HfO2 thin-film based on an elementary cost analysis.

In this work, the development and application of distributed economic model predictive control (DEMPC) methodologies to a catalytic reactor is considered. Specifically, two DEMPC methodologies are designed for sequential and iterative implementation, respec- tively. The DEMPC architectures are evaluated on the basis of the closed-loop performance and on-line computation time requirements compared to a centralized EMPC approach. For the catalytic reactor considered, DEMPC proves to be a viable option as it is able to give similar closed-loop performance while reducing the on-line computation time requirements relative to a centralized EMPC strategy.

Steam methane reforming (SMR) is the most widely used hydrogen (H2) production method,converting natural gas and steam into H2 and carbon dioxide (CO2). SMR is a well-studied process widely used in industrial applications to produce hydrogen, where fossil fuels are burned to provide heat for the endothermic reforming reactions, ultimately contributing to the production of greenhouse gas emissions. To mitigate emissions stemming from heating, a proposed solution involves utilizing an electrically-heated steam methane reformer process. Moreover, conventional SMR uses a packed bed catalyst and is heated by surrounding furnace; however, an electrified SMR employs a washcoated catalyst, is resistively-heated through the outer wall of the reactor, and loses heat to the surroundings. To further the foundational research conducted on electrically-heated reformers, this thesis examines the gas-phase products from an electrified reformer at UCLA, models it on process simulators, and scales up the proposed Aspen Plus model for industrial hydrogen production. The model simulates a plant with industrial level production rate and implements a ii reformer, two shift reactors, pressure swing adsorption (PSA) for separation, and a heat exchange network. The PSA process is modeled on Aspen Adsorption software, and the overall modeling process developed in Aspen Plus is discussed in detail. Lastly, a sensitivity analysis is performed on the entire process to determine the most energy-efficient conditions

Greenhouse gas emissions from industry and transportation contribute significantly to climate change and its associated adverse impacts on the environment and economy. With the increase in electricity supply from clean energy sources, electrochemical reduction of carbon dioxide (CO2) has received increasing attention; however, a first-principles model for electrochemical CO2 reduction has not been fully developed because of the complexity of its reaction mechanism. At this point, data driven methods and machine learning (ML) can be used to model the process. At UCLA, we have constructed an experimental rotating cylinder electrode cell (RCE) setup to develop a deeper understanding of the process, and design scale-up strategies. The experimental equipment is digitalized on a computer interface using Smart Manufacturing principles, legacy sensors (such as Gas Chromatogram (GC)) are automated, and voluminous steady state and dynamic data sets are generated. Leveraging these datasets and data analytics, machine learning and hybrid models are built using machine learning methods, such as Support Vector Machines (SVR), and artificial neural networks (ANN) and recurrent neural networks (RNN), that are capable of capturing nonlinearities and time dependencies. These models are used to optimize the most profitable setpoints and are used in single-input single-output (SISO) and multi-input multi-output (MIMO) feedback control schemes using multiple proportional integral (PI) controllers and model predictive control. This study proposes approaches for experimental implementation such as incorporating delayed GC feedback into control loops, training dynamic ML models with dead times, data variability and noise, and linearizing the RNN models using Koopman operators for fast real-time optimization.

The electrochemical CO2 reduction process has potential to use other chemical processes as CO2 source while sustainably producing other useful chemicals. In addition to the RCE setup, an experimental electrically-heated steam methane reforming (SMR) setup for hydrogen production from natural gas is digitalized, modeled, and controlled. A lumped parameter approach is used for fast calculations, and an extended Luenberger observer is used to compensate for missing and delayed sensor feedback. Finally, a model predictive controller using the estimation scheme is experimentally implemented to control hydrogen flow rates by manipulating the current, showing much faster response compared to PI control.

Plasma-enhanced atomic layer deposition (PEALD) has demonstrated its superiority at coatingultra-conformal high dielectric thin-films, which are essential to the fin field-effect transistors (FinFETs) as well as the advanced 3D V-NAND (vertical Not-AND) flash memory cells. Despite the growing research interest, the exploration of the optimal operation policies for PEALD remains a complicated and expensive task. Our previous work has constructed a comprehensive 3D multiscale computational fluid dynamics (CFD) model for the PEALD process and demonstrated its potential to enhance the understanding of the process. Nevertheless, the limitation of computational resources and the relatively long computation time restrict the efficient exploration of the operating space and the optimal operating strategy. Thus, in this work, we apply a 2D axisymmetric reduction of the previous 3D model of PEALD reactors with and without the showerhead design. Furthermore, a data-driven model is derived based on a recurrent neural network (RNN) for process characterization. The developed integrated data-driven model is demonstrated to accurately characterize the key aspects of the deposition process as well as the gas-phase transport profile while maintaining computational efficiency. The derived data-driven model is further validated with the results from a full 3D multiscale CFD model to evaluate model discrepancy. Using the data-driven model, an operational strategy database is generated, from which the optimal operating conditions can be determined for the deposition of HfO2 thin-film based on an elementary cost analysis.

This work focuses on applying machine learning modeling on predictive control of nonlinear processes with noisy data. We use long short-term memory (LSTM) networks with training data from sensor measurements corrupted by two types of noise: Gaussian and non-Gaussian noise, to train the process model that will be used in a model predictive controller (MPC). We first discuss the LSTM training with noisy data following a Gaussian distribution, and demonstrate that the standard LSTM network is capable of capturing the underlying process dynamic behavior by reducing the impact of noise. Subsequently, given that the standard LSTM performs poorly on a noisy dataset from industrial operation (i.e., non-Gaussian noisy data), we propose an LSTM network using Monte Carlo dropout method to reduce the overfitting to noisy data. Furthermore, an LSTM network using co-teaching training method is proposed to further improve its approximation performance when noise-free data from a process model capturing the nominal process state evolution is available. A chemical process example is used throughout the study to illustrate the application of the proposed modeling approaches and demonstrate their open- and closed-loop performance under a Lyapunov-based model predictive controller with state measurements corrupted by industrial noise.