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

Since the structures of crystals/moleculesare often non-Euclideandata in real space, graph neural networks (GNNs) are regarded as themost prospective approach for their capacity to represent materialsby graph-based inputs and have emerged as an efficient and powerfultool in accelerating the discovery of new materials. Here, we proposea self-learning-input GNN framework, named self-learning-input GNN(SLI-GNN), to uniformly predict the properties for both crystals andmolecules, in which we design a dynamic embedding layer to self-updatethe input features along with the iteration of the neural networkand introduce the Infomax mechanism to maximize the average mutualinformation between the local features and the global features. OurSLI-GNN can reach ideal prediction accuracy with fewer inputs andmore message passing neural network (MPNN) layers. The model evaluationson the Materials Project dataset and QM9 dataset verify that the overallperformance of our SLI-GNN is comparable to that of other previouslyreported GNNs. Thus, our SLI-GNN framework presents excellent performancein material property prediction, which is thereby promising for acceleratingthe discovery of new materials.

Automated summary

We propose a self-learning-input graph neural network framework, called SLI-GNN, to predict the properties of both crystals and molecules. By using a dynamic embedding layer and the Infomax mechanism, the input features are dynamically updated and the average mutual information between local and global features is maximized. Experimental results show that our SLI-GNN achieves comparable performance to other GNNs in material property prediction, indicating promising potential for accelerating new material discovery.