A graph neural network model with local environment pooling for predicting adsorption energies

Adsorption energy is an important descriptor of catalytic activity in modelling heterogeneous catalysis and is used to guide novel catalyst discovery. Previously, graph neural networks (GNNs) with global pooling have been applied to predict adsorption energy on different materials. However, adsorption energy is determined largely by the atoms near the adsorbate, and thus using local pooling methods that focus on this local chemical environment should enhance catalytic predictions. This study demonstrates the use of local environment pooling instead of the global pooling in conjunction with GNNs to predict adsorption energy. Based on the neural message passing with edge updates network and DimeNet++, we achieved mean absolute errors (MAEs) of 0.096 and 0.073 eV in predicting CO and H adsorption energies, respectively, on transition metal catalyst surfaces. Notably, these values surpass the performance of previously reported state-of-the-art machine learning models that employed the labelled site crystal graph convolutional neural network.

Automated summary

This study demonstrates the use of local environment pooling instead of global pooling in conjunction with graph neural networks to predict adsorption energy. We achieved mean absolute errors of 0.096 and 0.073 eV in predicting CO and H adsorption energies, respectively, on transition metal catalyst surfaces, surpassing the performance of previously reported state-of-the-art machine learning models.