Development of optimal control methods for unseeded batch cooling crystallization: Combination of first-principle and machine-learning approaches
Author(s)
Kim, Youngjo
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This thesis reports a framework to control the mean volume size and mass of paracetamol crystals in ethanolic solution for batch cooling crystallization. This framework utilizes the Markov state model (MSM) and dynamic programming (DP) approaches based on simulation results by a population balance model (PBM) model to obtain the optimal control policy for crystallization.
Since the MSM is an empirical model, a training data set is required, and numerous data points are needed to improve the model accuracy. To reduce the experimental attempts to establish the MSM, PBM simulation results are employed instead of experimental data. The PBM includes kinetic models for primary nucleation, secondary nucleation, crystal growth, and crystal dissolution. Crystallization experiments were carried out with temperature cycling, and kinetic parameters of the PBM were estimated and validated using the experimental data set. Since the PBM can predict the crystallization processes, this model generates data points to train the MSM. The trained MSM and DP approach optimizes the control policy to obtain desired crystal properties. The policies are tested by the PBM simulation and open-loop control experiments. However, it is challenging to get desired crystal properties using the open-loop control scheme due to the thermal response delay in the experimental system. Also, nucleation time is stochastic.
In addition, a feedback control scheme with an updated optimal control policy was employed to obtain the desired crystals. Since the process analytical technology (PAT) measurements, such as the focused beam reflectance measurement (FBRM), differ from the reduced-order states in the MSM, a model was built to convert the measurements into reduced-order states. A shallow neural network (SNN) model is developed for the data translation, and the crystallization system employs this model to monitor the solution status during the feedback control. The feedback control automatically manipulates the temperature profile to obtain crystals with the desired characteristics, and the control processes are completed when the system condition meets the control criteria. This thesis combines the first-principle model with a machine learning approach to demonstrate a process to control the mean volume size and crystal mass in unseeded batch cooling crystallization.
Sponsor
Date
2021-05-11
Extent
Resource Type
Text
Resource Subtype
Dissertation