Samsung Advanced Institute of Technology

In this study, we present a data-driven method for uncertainty-aware prediction of chemical reaction yields. A chemical reaction is represented as a set of molecular graphs for reactants and products. The predictive distribution over the yield is modeled as a graph neural network adapted to directly process the graph representations with permutation invariance. Uncertainty-aware learning and inference are applied for the model to make an accurate prediction as well as to abstain if the model is unconfident in the prediction. We demonstrate the effectiveness of the proposed method on benchmark datasets.