Chemical Reaction Mining

Many new and exciting chemicals have been discovered. However, the reactions to produce them often produce low reaction yields or require expensive intermediate reactants. These chemicals cannot be used because they are too difficult to create at the large scales required for industrial use. In order to remedy this problem, finding new reactions to produce desired products with improved yields or scalability is very desirable. To tackle this problem, I propose using transformer neural network architectures to augment molecule representations for predicting molecule reactions. In this project, I frame the problem as an information retrieval task, and I achieve a Hits@1 value of 50%. Further, I try to predict the reaction yield, although this does not prove effective using the transformer architecture.