Automated knowledge discovery for reaction engineering
File(s)
Author(s)
de Carvalho Servia, Miguel Ángel
Type
Thesis
Abstract
Accurate, interpretable kinetic models are essential for designing, optimizing, and controlling chemical processes. Traditional mechanistic modeling demands domain expertise, whereas purely data-driven approaches often lack transparency and physical consistency. This thesis presents automated knowledge discovery frameworks that address this by combining symbolic regression, information-theoretic model selection, and physicochemical constraints. Symbolic regression denotes machine learning methods that infer expressions directly from data, rather than assuming a fixed model form, aiming to extract plausible kinetic models with minimal expert intervention.
Chapter 3 examines model selection. A range of information criteria -- including Akaike, Bayesian, generalizable, and Hannan-Quinn -- are assessed to discriminate among kinetic models under noise, limited data, and data richness. A case study illustrates their practical behavior, and model-based design of experiments for model discrimination is introduced.
Chapter 4 introduces two symbolic regression frameworks, ADoK-S and ADoK-W (Automated Discovery of Kinetics, strong and weak formulations), which automate the discovery of rate laws from rate or concentration data. Both rely on genetic programming for expression generation and information-criterion-based refinement, and recover ground-truth models from sparse, noisy datasets.
Chapter 5 extends symbolic regression to mechanism discovery via SiMBA (Simplest Mechanism Builder Algorithm), a four-stage pipeline for mechanism generation, model translation, parameter estimation, and information-criterion-based model comparison. Parallelized backtracking and matrix representations enable efficient exploration of chemically feasible networks, as demonstrated on aldol condensation and fructose dehydration.
Finally, Chapter 6 proposes PI-ADoK (Physics-Informed Automated Discovery of Kinetics), which embeds physicochemical constraints into symbolic regression and incorporates uncertainty quantification via Metropolis-Hastings. Across several catalytic benchmarks, PI-ADoK improves model fidelity, robustness, and data efficiency. Collectively, these contributions advance automated kinetic model discovery for reaction engineering and process systems applications.
Chapter 3 examines model selection. A range of information criteria -- including Akaike, Bayesian, generalizable, and Hannan-Quinn -- are assessed to discriminate among kinetic models under noise, limited data, and data richness. A case study illustrates their practical behavior, and model-based design of experiments for model discrimination is introduced.
Chapter 4 introduces two symbolic regression frameworks, ADoK-S and ADoK-W (Automated Discovery of Kinetics, strong and weak formulations), which automate the discovery of rate laws from rate or concentration data. Both rely on genetic programming for expression generation and information-criterion-based refinement, and recover ground-truth models from sparse, noisy datasets.
Chapter 5 extends symbolic regression to mechanism discovery via SiMBA (Simplest Mechanism Builder Algorithm), a four-stage pipeline for mechanism generation, model translation, parameter estimation, and information-criterion-based model comparison. Parallelized backtracking and matrix representations enable efficient exploration of chemically feasible networks, as demonstrated on aldol condensation and fructose dehydration.
Finally, Chapter 6 proposes PI-ADoK (Physics-Informed Automated Discovery of Kinetics), which embeds physicochemical constraints into symbolic regression and incorporates uncertainty quantification via Metropolis-Hastings. Across several catalytic benchmarks, PI-ADoK improves model fidelity, robustness, and data efficiency. Collectively, these contributions advance automated kinetic model discovery for reaction engineering and process systems applications.
Version
Open Access
Date Issued
2025-09-17
Date Awarded
01/01/2026
License URL
Advisor
Hellgardt, Klaus
Hii, King Kuok (Mimi)
del Rio Chanona, Ehecatl Antonio
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/S023232/1
Publisher Department
Department of Chemical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)