ADVANCING MULTIPHASE PROCESS DEVELOPMENT WITH HYBRID PHYSICS-BASED – MACHINE LEARNING MODELS

Multiphase reaction systems are widely used in the chemical industry for the synthesis of pharmaceuticals, fine chemicals, and bulk intermediates. However, scaling up such systems is challenging due to the coupled effects of mass transfer and intrinsic reaction kinetics, both of which can control overall process behaviour. To develop predictive and generalizable models, it is essential to capture both reaction kinetics and the complex transport phenomena that influence system performance. This thesis presents a hybrid modelling framework for reactive multiphase processes, demonstrated on a liquid-liquid nitration system, and developed using experimental data guided by strategies for informative experiment selection. Understanding hydrodynamic behaviour is critical in such systems, as it determines interfacial area and mass transfer characteristics, both of which impact reaction performance. A workflow is developed to extract hydrodynamic descriptors and estimate interfacial area in a liquid–liquid system in a microreactor. High-speed imaging and computer vision techniques are applied to quantify flow characteristics, and a machine learning model is trained to predict the specific surface area from operating conditions. In addition, accurate estimation of kinetic parameters in multiphase systems often requires extensive experimentation, as parameters may be correlated and not all operating conditions contribute equally to model calibration. To address this, a model-based design of experiments (MBDoE) approach is developed to identify the most informative experimental conditions based on the structure of the mechanistic model. Building on these elements, a hybrid modelling framework is proposed by integrating a mechanistic kinetic model with a machine learning model for predicting the mass transfer coefficient, alongside the specific surface area model. The kinetic model is experimentally calibrated using MBDoE on an automated continuous flow platform, which enables the targeted acquisition of high-quality data. An active learning strategy is also implemented on the same platform to iteratively suggest new experiments that inform the mass transfer predictions and improve overall model accuracy. The use of physically meaningful inputs to the machine learning component contributes to the model’s ability to generalize and accurately capture system behaviour across a wide range of operating conditions. Overall, the thesis brings together experimental, process modelling, and machine learning components with the shared goal of streamlining the development of reliable hybrid models for complex multiphase systems.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Katsarou, Anna

Advisor dc:contributor.advisor

• Lapkin, Alexei

• Machine learning
• Design of experiments
• Reaction kinetics

Licence

• https://www.repository.cam.ac.uk/bitstreams/0b751d7a-3ffd-49ab-93ad-b0f08ed923f0/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.121350

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/389485