AI3SD intern project: Reaction condition optimisation

Abstract

Optimisation problems are common throughout chemistry, for example optimising the yield of a chemical reaction. Since chemical reactions can take hours or days to complete, it is infeasible
to search through the entire set of possible conditions. Instead, a small subset of the search space is tested, with Design of Experiments (DOE) and Generalised Subset Design being two common techniques used to determine such a subset. Bayesian optimisation, an iterative global optimisation algorithm, has found success in hyperparameter tuning for machine learning models, where researchers face a similar issue of long model evaluation times. As a result, it has been recently applied to problems in chemistry, including the development of the EDBO (Experimental Design for Baysian Optimisation) optimiser, which was used throughout the project. We aimed at verifying the results obtained from the EDBO paper, and then testing the robustness of the optimiser by applying it to other datasets, including one outside the scope of its original domain of reaction yield optimisation.

NB: This report was formalised into a journal paper which has been published in the Journal of Cheminformatics: Robustness under parameter and problem domain alterations of Bayesian optimization methods for chemical reactions. This report serves as a preprint of that paper.

More information

Identifiers

ORCID for Stephen Gow: orcid.org/0000-0003-0121-1697

ORCID for Samantha Kanza: orcid.org/0000-0002-4831-9489

ORCID for Jeremy G. Frey: orcid.org/0000-0003-0842-4302

Catalogue record

Export record

Altmetrics