Doctor of Philosophy
Chemical and Biochemical Engineering
Driven by the increasing demands of producing consistent and high-quality crystals for high value-added products such as pharmaceutical ingredients, the operation and design of a crystallization process have phased from an empirical trial-and-error approach to the modern frameworks powered by the online process analytical technologies (PATs) and model-based process optimization techniques.
The one-dimensional crystal size distribution (CSD) measured by the well-established PATs is inadequate due to the missing particle morphology information. A major contribution of this thesis is to develop an image analysis-based PAT powered by the deep learning image processing techniques, whose accuracy and functionality outperformed the traditional PATs and other image analysis techniques. The PAT was deployed to monitor and study the slurry mixture of glass beads and catalyst particles as well as a taurine-water batch crystallization process. The results confirmed the superb accuracy of two-dimensional size and shape characterization in a challengingly high solids concentration. The classification capability enabled unparalleled functionalities including quantification of agglomeration level and characterization of different polymorphs based on their distinct appearances. A computerized crystallization platform was built with the developed PAT, which could automate the time-consuming experiments for determining the metastable zone width (MSZW) and induction time of a crystallization system. The application of the PAT revealed the potential to simplify and speed up the research and development stage of a crystallization process.
The rich two-dimensional crystal size and shape information provided by our PAT enabled more descriptive multi-dimensional modeling for the better prediction of the crystallization process. The novel population array (PA) solver developed in this thesis could solve the multi-dimensional crystallization population balance equation (PBE) more computationally efficient than the existing discretization-based numerical methods without compromising the accuracy. The PA solver could accurately model the complex phenomena including agglomeration, breakage, and size-dependent growth. The efficient computation enables solving the complex multi-dimensional PBE for crystal morphology modeling. The combination of the innovative PAT and modeling technique is a significant contribution to the crystallization field that enables better understanding and more effective control of a crystallization process.
Summary for Lay Audience
Many food, drug, and industrial products are in crystalline form. The size, shape, and many other attributes of the crystals will affect the process efficiency and product quality, which is especially critical for the pharmaceutical industry. Therefore, process analytical technologies (PATs) were developed to monitor the process variables and ensure product quality.
An efficient PAT instrument for the size and shape characterization, which was powered by the latest artificial intelligence algorithms was developed. The accuracy of the proposed PAT was validated by comparing the measurement results with the established particle sizing methodologies. The PAT was proven not only accurate but could also differentiate the type of the crystals and perform multi-dimensional size measurement, which was very challenging for the existing instruments. The crystal size and shape data were used to build a dynamic model to predict and optimize the product quality of a crystallization process.
The crystallization modeling helps to design optimal crystallization processes. A numerical solver was proposed to efficiently solve the mathematical model of a crystallization process. The proposed solver was tested under various scenarios. The accuracy and computational efficiency have improved compared to the conventional approaches.
The PAT and the modeling techniques were combined in the development of an automated crystallization experiment platform, which aimed to automate the repetitive and time-consuming crystallization experiments. The homemade instruments enabled automated data acquisition and process manipulation. With the automated platform, the research and development of a crystallization process could be significantly simplified.
Wu, Yuanyi, "A Study of The Deep Learning-based Monitoring and Efficient Numerical Modeling Methodologies for Crystallization Processes" (2021). Electronic Thesis and Dissertation Repository. 7813.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.