(6cs) Process Systems Engineering and Artificial Intelligence for Advanced Manufacturing: Including Applications to Biopharmaceuticals

Achieving high product quality has become more challenging as the complexity and scale of manufacturing rapidly increase. The U.S. Food and Drug Administration (FDA), for instance, has been shifting the pharmaceutical industry away from the traditional quality by testing (QbT) approach, which involves univariate analyses of product quality by trial-and-error and heavily relies on heuristics. Instead, more systematic and model-based methodologies such as quality by design (QbD) and quality by control (QbC) are gradually being adopted. My research interests revolve around addressing core challenges for QbD and QbC such as efficient processing of manufacturing and R&D data, developing reliable models, and designing optimal processes and control strategies with the help of process systems engineering (PSE) and artificial intelligence (AI).

PSE is a chemical engineering discipline that focuses on developing systematic and computer-assisted approaches to the design, operation, optimization, and control of processes. At its core, PSE is about systems thinking, which considers a system, its interacting constituents, and competing objectives. QbD and QbC could benefit from both the technological advances in PSE—progress in model predictive control (MPC), multi-objective optimization, global optimization, nonlinear control and optimization, etc.—and its unique systems perspective.

Due to the significant role that computing plays in supporting PSE, PSE also has a long history of applying AI techniques. The field of AI research was born more than sixty years ago and early PSE adoptions date back to the 1980s (e.g., expert systems). Recent advances in computing technologies (power, storage, and infrastructure) have led to rapid developments in certain AI fields such as machine learning and natural language processing (NLP), providing new tools for PSE to adopt in the age of big data. The abundance of data in manufacturing creates both opportunities and challenges: On the one hand, the scale of process data is enormous and new data are created every day. On the other hand, much of the data are unorganized, not machine-readable, prone to human error, or contain proprietary information. These limitations make it difficult to generate useful information from the data algorithmically. The challenges are especially pronounced in certain manufacturing fields such as biopharmaceuticals where large-scale computing has yet to be widely adopted.

I come from a diverse and unusual research background: My doctoral training from Columbia University with Prof. Venkat Venkatasubramanian and Prof. Garud Iyengar focused on solving sociotechnical problems with process control, risk management, data science, and game theory. I have collaborated with leading experts and executives in chemical engineering, computer science, operations research, finance, economics, and public health. Such diverse research experience gives me profound understanding in risk management and decision-making as well as the capacity to conduct high-quality interdisciplinary research. My current research with Prof. Babatunde Ogunnaike and Prof. Kelvin Lee at the University of Delaware involves a close collaboration with a major pharmaceutical company and the application of system identification, process design, design of experiments (DoE), and multiscale modeling to biopharmaceutical processes for the improvement of production and product quality. Our hybrid model (knowledge-based and data-driven) achieves high prediction accuracy; its modular structure can be easily updated with new process inputs such as experimental conditions, media/feed compositions, and sampling/feeding strategies. I have also optimized the computer codes, making them run sixty times faster than the previous version. The deliverable for my first six-month collaboration was a fully functional user interface that can predict production and product quality and also solve process variables to meet certain production and product quality goals thus achieving QbD.

My understanding of the key challenges in advanced manufacturing is as follows. First, one of the most important tasks central to achieving the objectives of QbD and QbC is the development of high-fidelity models that can reliably predict production and product quality. I will take advantage of the computing breakthroughs (e.g., deep learning, GPU computing, and NLP) to analyze (structured and unstructured) process data and extract information regarding process dynamics for the purpose of building data-driven models. Furthermore, I will use global optimization tools such as genetic algorithm, simulated annealing, and particle swarm optimization to design and discover optimal processes and control strategies. This will be a natural extension of my postdoctoral work.

Second, the positive benefits of rapid technological developments sometimes come with negative unintended consequences to society. Unintended consequences, by definition, are difficult to detect. For example, efficient light bulbs could increase overall energy consumption due to the low cost. I will apply PSE techniques (e.g., fault detection and diagnosis), risk management techniques (e.g., failure mode and effects analysis), and AI techniques (e.g., applied ontology and various search algorithms) to algorithmically identify hazards and reduce negative unintended consequences of manufacturing.

Last but not the least, decision-making process in manufacturing can benefit from the “human-in-the-loop” computing—algorithms generate timely recommendations based on models and data while human decision-makers evaluate, accept, modify, or reject such recommendations. This arrangement, if done right, would benefit both human decision-makers and algorithms. On the one hand, algorithms are quick, exhaustive, and objective. On the other hand, human judgment and feedback in return can serve as supervised learning to improve algorithms. In addition to human–machine interactions, according to my doctoral research, the collective intelligence of individual human decision-makers can be potentially enhanced through certain types of information sharing. Developing mechanisms that integrate both human and machine intelligence requires insight beyond engineering. My prior experience with social science gives me an edge on creating advanced decision-making framework through “human-in-the-loop” computing.

Given sufficient preparation, I can teach any chemical engineering course. That said, my first priority is teaching courses related to my research background, such as process dynamics and control, process safety, design project, simulation, and mathematics (differential equation, numerical analysis, linear algebra, etc.). I also plan to introduce new content about relevant and current technological advances to enhance students’ understanding and motivate them to learn. Examples of topics include but are not limited to MPC, risk management, and AI.

In addition to teaching the core curriculum, I am also interested in designing technical electives on topics such as computing, project management, data science, and optimization, either as standalone courses, or packaged together as a single course that covers the basics of each topic. The purpose of these electives is to guide students to use tools and techniques outside chemical engineering—code documentation, code optimization, coding style, version control, project planning, project monitoring, SQL, deep learning, linear programming, convex optimization, etc.—to conduct research more efficiently.