Automated Optimization and Control of Modular Chemical Systems

Continuous manufacturing has been widely used in many industries such as food, oil, and gas. The pharmaceutical sector is now transitioning from batch to continuous manufacturing as well, to achieve potential benefits such as improved responses to demand changes and reduction of supply chain disruptions, production time, drug costs, waste, and product quality variations. Continuous pharmaceutical manufacturing in compact modular systems provides additional flexibility in where and when the drug is produced. The transition to continuous operation raises new control challenges to meeting specifications in critical quality attributes (CQAs). Process models, facilitated by the widespread adoption of process analytical technology for on-line CQA measurements, enable the concise formulation of existing process understanding and promote realtime decision making and retrospective analyses of the process of interest. First-principles or data-driven models can be developed depending on the degree of process understanding and the purpose of the model construction. Models can be used in process optimization to improve experimental design and manufacturing practices. Models are also used in model-based control to handle variations which would lead to reduced product quality. Model-based control offers opportunities for bringing processes a step closer towards full automation. This thesis employs, enhances, and develops system engineering tools to address challenges associated with automating optimization and control solutions for modular chemical systems. First, the thesis presents first-principles mathematical descriptions for common modules in a modular chemical system. An approach is developed for the derivation of linear input–output (step-response) models that reduce model–plant mismatch. The proposed step-response models allow for the successful implementation of a variation of linear model predictive control (MPC), known as quadratic dynamic matrix control. Dynamic optimization for startup based on a first-principles plant-wide model is formulated and solved for a virtual plant for the upstream synthesis of atropine. Then the thesis presents a methodology for designing stabilizing dynamic state feedback controllers and observers with guaranteed properties for dynamic artificial neural network (DANN) models using matrix inequalities. Assuming a known DANN structure describing a system, a more conservative representation known as a diagonal norm-bounded linear differential inclusion is employed to derive sufficient criteria for estimation and control in the form of linear or bilinear matrix inequality problems using quadratic Lyapunov functions. A computational case study demonstrates the applicability of the method in a realistic multiple-input multiple-output pH control problem. Lastly, the thesis proposes a strategy for constructing nonlinear interpretable input–output models for modular chemical systems. Polynomial nonlinear-autoregressive-with-exogenous-inputs models are identified using a machine learning algorithm that promotes variable selection and grouping of correlated variables, and results in a sparse representation. The models are incorporated into a nonlinear MPC algorithm implemented in the JuMP algebraic modeling language, which results in very fast computational times. Computational case studies for two different types of chemical reactors demonstrate the applicability of the methodology in processes that commonly appear in modular chemical systems.Ph.D