Prediction of Reaction Performance by Machine Learning Using Streamlined Features: NMR Chemical Shifts as Familiar Descriptors

Abstract

Machine learning (ML) has quickly emerged in synthetic organic chemistry to predict reaction outcomes such as yields and stereoselectivities. Notably, recent applications of the ML approach showed powerful performance in solving various chemical problems. However, the requirement of numerous descriptors and large datasets hampers the general use by non-specialists. In this study, simple ML models were developed by utilizing easily available ¹³C-NMR chemical shifts of the substrates as familiar descriptors to predict the site-selectivity of geminal chlorofluorination of unsymmetrical 1,2-dicarbonyl compounds. We identified that the feed-forward neural network (FNN) model provides higher accuracy compared to other algorithms. Then, better prediction performance was acquired through streamlined models using minimal, only empirically relevant descriptors.