Predicting Reactions with Controlled Errors
Given a group of molecules and a specification of reaction conditions, do chemical reactions occur? If so, what products are produced, and what is their relative abundance? This problem pervades chemistry, with applications in environmental science (e.g., the degradation of pollutants), molecular sensing (e.g., interpreting the results of tandem mass spectrometry), chemical synthesis (e.g., finding efficient ways to synthesize drug molecules), and energy-efficiency (characterizing molecular combustion). Such problems are usually addressed by nearly exhaustive experimental and/or computational characterization techniques, both of which are extremely costly in terms of time, money, and human resources. We aim to use a data-science approach, so that these previous experimental/computational works can be leveraged to make predictions of likely chemical reactive pathways. One important innovation is to make these predictions in a controlled way, with error estimates so that future experimental/computational studies can be focused towards reactions where the model is highly uncertain, and redundant work can be avoided. Other innovations include using reactivity indicators to construct rich molecular representations and representing chemical reaction networks as hypergraphs.
View Full Project DescriptionPaul Ayers
Sorbonne Université
Physics
Artificial Intelligence; Quantum Science
McMaster University
Globalink Research Award