CFD-DNS analysis and machine learning predictions of dispersion dynamics in complex multiphase flows in static mixers
File(s)
Author(s)
Valdes, Juan Pablo
Type
Thesis
Abstract
L-L mixing has attracted substantial academic and industrial interest in recent decades due to its applications across various industries and its numerous advantages over traditional mixing devices. The inherent complexities of L-L flows present challenges for developing design, scale-up, and process modelling frameworks. Despite the extensive body of literature, current understanding of the governing fundamental physics remains limited and reliant on several modelling assumptions, calling for further research. These challenges have encouraged the development of computational tools aiming to provide efficient and reliable solutions for industrial applications. This PhD conducts a two-fold computational investigation. At its core, high-fidelity DNS is deployed, leveraging a state-of-the-art interface tracking algorithm that unlocks unprecedented access to a wealth of information on the interfacial dynamics under industrially relevant conditions.
Secondly, a DL framework is implemented to perform cost-effective dispersion performance predictions. Additionally, an early-stage active-learning ML surrogate model is introduced and tested as a complementary tool for mixer design. A general understanding of the governing deformation and breakup mechanisms in simplified scenarios is elucidated via DNS and extrapolated to more intricate cases. Insights on key surfactant-laden dynamics are examined, including the interrelationships between Marangoni stresses and local hydrodynamics, revealing a dual effect on deformation and breakup. In addition, trends in dispersion performance are identified. Exploratory DNS analyses were conducted on the impact of mixer geometry on dispersion performance. The DL platform proposed exhibited satisfactory performance in capturing primary behavioural patterns per case, despite the limited training data. The fully connected LSTM network excelled in generalization and performance by effectively managing noise and information flow, capturing long-term dependencies better. An early-stage surrogate model framework was also tested as a means to inexpensively predict internal flow and mixing metrics as functions of geometrical parameters, showing promising potential and laying the groundwork for future enhancements.
Secondly, a DL framework is implemented to perform cost-effective dispersion performance predictions. Additionally, an early-stage active-learning ML surrogate model is introduced and tested as a complementary tool for mixer design. A general understanding of the governing deformation and breakup mechanisms in simplified scenarios is elucidated via DNS and extrapolated to more intricate cases. Insights on key surfactant-laden dynamics are examined, including the interrelationships between Marangoni stresses and local hydrodynamics, revealing a dual effect on deformation and breakup. In addition, trends in dispersion performance are identified. Exploratory DNS analyses were conducted on the impact of mixer geometry on dispersion performance. The DL platform proposed exhibited satisfactory performance in capturing primary behavioural patterns per case, despite the limited training data. The fully connected LSTM network excelled in generalization and performance by effectively managing noise and information flow, capturing long-term dependencies better. An early-stage surrogate model framework was also tested as a means to inexpensively predict internal flow and mixing metrics as functions of geometrical parameters, showing promising potential and laying the groundwork for future enhancements.
Version
Open Access
Date Issued
2024-07-15
Date Awarded
01/12/2024
License URL
Advisor
Matar, Omar
Kahouadji, Lyes
Cheng, Sibo
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/T000414/1
Publisher Department
Chemical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)