Samsung Advanced Institute of Technology

In synthetic planning, it is important to find feasible reaction conditions that make a chemical reaction work as intended.
Recent research attempts based on machine learning have proven to be effective in recommending reaction elements for some specific categories regarding critical chemical context and operating conditions. However, existing methods can only make a single prediction per reaction and does not directly provide a full specification of reaction elements as the prediction. Thus, their achievable performance is limited in practice. In this study, we propose a generative modeling approach to predict multiple different reaction conditions for a chemical reaction, each of which fully specifies critical reaction elements such that these can be directly used as an actionable reaction condition.
We formulate the problem of reaction condition prediction as sampling from a generative distribution.
 We model the distribution by introducing a variational autoencoder augmented with a graph neural network and learn it from a reaction dataset. For a query reaction, multiple predictions of reaction conditions can be obtained by repeated sampling from the distribution.
Through experimental investigation on the reaction datasets of four major cross-coupling reaction types, we demonstrate that the proposed method significantly outperforms the existing methods in retrieving ground-truth reaction conditions.