Modelling Organic Reactivity with Machine Learning and Quantum Chemistry

Abstract

The computational design of synthetic methodology remains a central ambition of computational chemistry, yet the cost of high-accuracy quantum-mechanical calculations limits how far rational design can scale. Density functional theory (DFT) has proven indispensable for mechanistic insight and reactivity prediction in individual transformations, but extending this approach to the thousands of conformations and reaction pathways encountered in real synthetic campaigns quickly becomes prohibitive.

This thesis demonstrates that machine learning (ML) models trained on inexpensive semi-empirical calculations can recover near-DFT accuracy while reducing computational cost by orders of magnitude, enabling predictive workflows that extend beyond analysis toward molecular design. A mechanistic study of a [1,2]-Wittig rearrangement first reveals the computational challenges associated with exhaustive conformational and transition-state exploration, identifying key bottlenecks that constrain traditional quantum-chemical approaches. These bottlenecks are then addressed through ML models capable of predicting activation free energies from semi-empirical quantum mechanical (SQM) descriptors, achieving near-DFT performance with substantially reduced computational expense. The scope of this strategy is further examined through attempts to learn deviations between SQM and DFT molecular geometries; while these models do not achieve the accuracy required for practical deployment, their performance highlights important distinctions between learning energetic and structural quantities and helps define the limits of the current approach. Finally, predictive models of reactivity are incorporated into a generative framework that proposes candidate molecules satisfying target activation barriers and refines them through active learning.

Together, these studies establish a scalable framework that links mechanistic understanding, predictive modelling, and molecular generation. By combining the interpretability of quantum chemistry with the efficiency of modern ML, this work outlines a practical route toward the computational design of new synthetic transformations.

Date of Award	22 Jul 2026
Original language	English
Awarding Institution	University of Bath
Supervisor	Matt Grayson (Supervisor) & Pranav Singh (Supervisor)