Graph based machine learning for accurate atom mapping of chemical reaction

Abstract

Atom mapping problem of chemical reaction has wide application in fields such as chemical reaction prediction, retrosynthesis, data construction, and mechanism simulation. In order to effectively map a large amount of chemical reaction data, fast information processing based on computational method is needed, rather than conventional experimental methods. In this thesis, we demonstrate that development of a machine learning model LocalMapper. LocalMapper train on organic chemical reactions based on graph artificial neural networks and attention artificial neural networks. LocalMapper perform accurate atomic mapping with only small amount of data by applying active learning method to reduce the amount of machine learning information processing and to make effective predictions for a very large amount of test chemical reactions. LocalMapper can provide the missing link between chemical reaction data and any further application with predicted accurate atom mapping.