Multi-objective optimization and model-based predictive control using state feedback linearization for crystallization

The ongoing Quality-by-Design paradigm shift in the pharmaceutical industry has sparked a new interest in exploring advanced process control techniques to aid the efficient manufacture of high value products. One important process in the manufacturing is crystallization, a key process in purification of active pharmaceutical ingredients (APIs). There has been little crystallization control research in the area of global input/output linearization, otherwise referred to as state-feedback linearization (SFL). The global linearization allows a nonlinear model to be linearized over the whole domain for which the model is valid and can be embedded into a model predictive controller (MPC). MPC allows the control of a process based on a model which captures the physical understanding and constraints, but a widely reported challenge with the SFL technique is the poor ability of explicitly handling the plant constraints, which is not ideal for a highly regulated production environment such as pharmaceutical manufacturing. Therefore, the first purpose of this research is to explore the use of SFL and how it can be applied to controlling batch and continuous MSMPR crystallization processes with the incorporation of plant constraints in the MPC (named SFL-Plant constraints). The contribution made from this research is the exploration of the SFL MPC technique with successful implementation of SFL-Plant constraints. The novelty in this method is that the technique builds on existing SFL-MPC frameworks to incorporate a nonlinear constraints routine which handles plant constraints. The technique is applied on numerous scenarios of batch and continuous mixed suspension mixed product removal (MSMPR) supersaturation control of paracetamol in water, both seeded and unseeded, which all show that the SFL-Plant constraints technique indeed produces feasible control over crystallization subject to constraints imposed by limitations such as heat transfer. The SFL-MPC with SFL-Plant constraints was applied to single-input single-output (SISO) and multiple-input multipleoutput (MIMO) systems, demonstrating consistent success across both schemes of control. It was also determined that the SFL-Plant constraints do increase the computational demand by 2 to 5 times that of the SFL when unconstrained. However, the difference in absolute time is not so significant, typically an MPC which acted on a system each minute required less than 5 seconds of computation time with inclusion of SFL-Plant constraints. This technique 5 presents the opportunity to use the SFL-MPC with real system constraints with little additional computation effort, where otherwise this may have not been possible. A further advancement in this research is the comparison between the SFL-MPC technique to an MPC with a data-driven model - AutoRegression model with eXogenous input (ARX) – which is widely used in industry. An ARX model was identified for batch supersaturation control using a batch crystallization model of paracetamol in isopropyl alcohol (IPA) in gPROMS Formulated Products as the plant, and an ARX model developed in an industrial software for advanced process control – PharmaMV. The ARX-MPC performance was compared with SFL-MPC performance and it was found that although the ARX-MPC performed well when controlling a process which operated around the point the ARX-MPC was initially identified, the capability of tracking the supersaturation profile deteriorated when larger setpoints were targeted. SFL-MPC, on the other hand, saw some deterioration in performance quantified through an increase in output tracking error, but remained robust at tracking a wide range of supersaturation targets, thus outperforming the ARX-MPC for trajectory tracking control. Finally, single-objective and multi-objective optimization of a batch crystallization process is investigated to build on the existing techniques. Two opportunities arose from the literature review. The first was the use of variable-time decision variables in optimization, as it appears all pre-existing crystallization optimization problems to determine the ideal crystallization temperature trajectory for maximising mean-size are constructed of piecewise-constant or piecewise-continuous temperature profiles with a fixed time step. In this research the timestep was added as a decision variable to the optimization problem for each piecewise continuous ramp in the crystallization temperature profile and the results showed that for the maximisation of mean crystal length in a 300-minute batch simulation, when using 10 temperature ramps each of variable length resulted in a 20% larger mean size than that of 10 temperature ramps, each over a fixed time length. The second opportunity was to compare the performance of global evolution based Nondominated Sorting Genetic Algorithm – II (NSGA-II) with a deterministic SQP optimization method and a further hybrid approach utilising first the NSGA-II and then the SQP algorithm. It was found that for batch crystallization optimization, it is possible for the SQP to converge a global solution, and the convergence can be guaranteed in the shortest time with little compromise using the hybrid 6 method if no information is known about the process. The NSGA-II alone required excessive time to reach a solution which is less refined. Finally, a multi-objective optimization problem is formed to assess the ability to gain insight into crystallization operation when there are multiple competing objectives such as maximising the number weighted mean size and minimizing the number weighted coefficient of variation in size. The insight gained from this is that it is more time efficient to perform single-objective optimization on each objective first and then initialize the multi-objective NSGA-II algorithm with the single-objective optimal profiles, because this results in a greatly refined solution in significantly less time than if the NSGA-II algorithm was to run without initialization, the results were approximately 20% better for both mean size and coefficient of variation in 10% of the time with initialization.