Data-driven optimisation and discovery for engineering systems
Author(s)
Savage, Tom
Type
Thesis
Abstract
This thesis explores the integration of machine learning, optimisation, and human expertise for scientific discovery and the improvement of engineering systems. Each chapter applies deductive reasoning through physics-based models, inductive reasoning through data-driven approaches, or a combination of the two, demonstrating how together they drive scientific progress. First, a methodology is established for designing novel chemical flow reactors by combining multi-fidelity Bayesian optimisation with computational fluid dynamics. This approach enables automated discovery of reactor geometries and validation of induced flow characteristics, improving performance by 60\% over conventional designs. Key features of optimal designs are identified and validated experimentally through 3D-printed reactors, demonstrating how mixing-enhancing vortical structures can be induced at previously unattainable conditions. Second, a human-algorithm collaborative Bayesian optimisation framework is presented that integrates domain expertise into data-driven decision-making. By exploiting the hypothesis that humans are more efficient at discrete rather than continuous choices, the framework enables experts to influence solution selection via Bayesian optimisation with minimal effort. Case studies across benchmark functions, reactor design, and bioprocess optimisation demonstrate that even partial expertise can accelerate convergence compared to standard methods, particularly for noisy, high-dimensional problems. Finally, in-silico improvements to Bayesian optimisation are explored using large language models (LLMs) as computational decision-makers. This approach maintains automation while demonstrating convergence improvements similar to those from expert intervention. By using LLMs to select between discrete candidate solutions, the methodology identifies patterns in optimisation trajectories and makes informed selections. This thesis establishes data-driven optimisation as a powerful framework for scientific discovery in engineering systems beyond the paradigms of modelling and optimisation. The practical improvements ranging from more efficient physical reactor designs to accelerated optimisation processes demonstrate how the methodologies presented in this thesis can advance scientific and engineering discovery.
Version
Open Access
Date Issued
2025-05-05
Date Awarded
01/11/2025
License URL
Advisor
del Rio Chanona, Antonio
Publisher Department
Department of Chemical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)