SLI-GNN: A Self-Learning-Input Graph Neural Network for Predicting Crystal and Molecular Properties

Since the structures of crystals/molecules are often non-Euclidean data in real space, graph neural networks (GNNs) are regarded as the most prospective approach for their capacity to represent materials by graph-based inputs and have emerged as an efficient and powerful tool in accelerating the discovery of new materials. Here, we propose a self-learning-input GNN framework, named self-learning-input GNN (SLI-GNN), to uniformly predict the properties for both crystals and molecules, in which we design a dynamic embedding layer to self-update the input features along with the iteration of the neural network and introduce the Infomax mechanism to maximize the average mutual information between the local features and the global features. Our SLI-GNN can reach ideal prediction accuracy with fewer inputs and more message passing neural network (MPNN) layers. The model evaluations on the Materials Project dataset and QM9 dataset verify that the overall performance of our SLI-GNN is comparable to that of other previously reported GNNs. Thus, our SLI-GNN framework presents excellent performance in material property prediction, which is thereby promising for accelerating the discovery of new materials.