An application of graph neural network in pollution abatement: Acceleration heterogeneous catalyst design

Abstract

The rapid development of Artificial Intelligence has made it possible for applying DFT to the design of complex catalysts. This paper investigates the effectiveness of Graph Neural Network (GNN) accelerated DFT calculation for designing dual atom catalysts (DACs). The GNN was trained with 231 sets of energy data pairs obtained from the DFT calculation, and the R2 on the training and test sets were 0.985, 0.952, respectively. Comparing the performance of traditional machine learning and GNN, the results show that GNN has irreplaceable advantages. We created a new descriptor to evaluate the overall performance of DACs, and obtained 7 outperforming DACs. Then, the effect of atomic properties on catalyst performance is explored, which reveals the contribution of individual properties to the predicted values and improves the interpretability of the models. Finally, we performed multiphase catalytic experiments to demonstrate the effectiveness of GNN accelerated DFT calculation in catalyst design. The successful application of GNN provides the necessary reference for designing complex catalyst systems.