Exploring Reaction Condition Space with Expert-Informed Representation Learning

The selection of appropriate reaction conditions is paramount in chemical synthesis and can greatly impact the success of experimental campaigns. Machine learning models for reaction outcome prediction have focused primarily on reactant representations, whereas the representation of reaction conditions (the chemical components defining the reaction environment, e.g. activators, bases or solvents) has received comparatively less attention. Here, we present a novel machine learning framework designed to generate chemically relevant representations for sets of conditions. We evaluate these representations using literature Amide Coupling High-Throughput Experimentation (HTE) datasets, testing whether distances in embedding space reflect local yield similarity, physical-property trends, and experimentally useful condition diversity. Our results demonstrate that the learnt representations capture reactivity-relevant trends and support downstream tasks including condition visualization, diversity-aware HTE plate design, and selection of alternative conditions.