Prediction of energies for reaction intermediates and transition states on catalyst surfaces using graph-based machine learning models

Abstract

Computational studies of heterogeneous catalysis processes depend on massive electronic structure calculations to obtain the energies of intermediates and transition states. To speed up this process, several machine-learning-based methods were proposed for the prediction of surface species energies. Here we developed a new method to represent all surface species with molecular graph, a data structure which is easy to read and extendable, but seldom utilized in catalysis studies. The molecular graph dataset consists of 315 C1/C2 surface intermediates and transition states on Rh(111), which are all possible intermediates in the complex reaction network of ethanol synthesis from syngas. Three recently proposed graph-based machine learning methods, namely graph convolutions, weave and graph neural network, were employed to train models and predict the energies from molecular graphs. Furthermore, two ensemble models combining the abovementioned models were built, using which the best RMSE and MAE reaches 0.19 and 0.15 eV, respectively. In addition, error of activation energies predicted with graph neural network was compared with that predicted using traditional BEP relations, and error of the prediction for surface intermediate energies and transition state energies were compared. Finally, possible directions of using the developed methods in extendable energy predictions were suggested and discussed.