Rapid enthalpy prediction of transition states using molecular graph convolutional network

The rapid development of machine learning has provided low-cost and high-throughput molecular energy prediction methods; however, the direct energy prediction of transition states (TSs) in radical reactions has received little attention, which can facilitate the rapid construction of potential energy surface (PES) and chemical reactivity inference. Here, we present a generic descriptor of TSs, molecules and radicals to realize the simultaneous prediction of their enthalpy of formation through the augmented molecular graph convolutional network (MGCN). It is proved that the network can rapidly (within milliseconds per prediction) and accurately predict the enthalpy of various types of samples, of which the mean absolute error (MAE) of the held-out test set is 1.44 kcal/mol (>50% samples less than 1.0 kcal/mol) with high R-2 (0.925-0.998) among all categories. The PES constructed from predictions is close to the calculated results by density functional theory (DFT) and conforms to the trend of chemical reactivity.