Machine learning has been used to accelerate the discovery in chemistry. Primarily, people use supervised learning to predict the properties of molecules and reactions to inform new developments. The prediction of molecular and reaction properties have therefore been at the center of machine learning research in chemistry. The key challenges exist in developing the most appropriate representations, choosing the best model architecture, and, in the long run, development of better-quality datasets.

This project focuses on develop models to represent chemical molecules and reactions, and use them to predict molecular properties like toxicity, solubility and reaction properties like yield, impurities. Different representations will be explored and physical or chemical knowledge can be incorporated to enhance the effectiveness of the models.