Adaptative Scaler-Moment Crystal Graph Attention Neural Network for Material Property Prediction

Graph neural networks have gained increasing popularity for predicting crystal material properties. However, significant dilemmas are involved in designing such models: (i) chemical information about crystals is difficult to capture, and (ii) a complex model is required to map the chemical space to the property space. In this study, we develop an adaptative scaler-moment crystal graph attention neural network (SM-CGANN) for predicting crystal material properties. The graph neural network is enhanced using scalar moment aggregation functions and attention mechanism, controlling chemical information exchange between the central atom and its neighbors. Graph pooling increases the information transmission rate by maximizing mutual information between the pooled and input graphs. In addition, we incorporate the multi-objective Bayesian optimization method to quickly find the best hyperparameters and network architecture, ensuring an adaptive balance between the prediction accuracy and spatial complexity of SM-CGANN. This method is superior to state-art-of models in terms of accuracy performance for different material properties in density function functional theory calculation datasets (Materials Project and Open Quantum Materials Database). Moreover, it provides highly accurate performance of end-user scenarios involving the classification of metal/nonmetal and high-/weak-magnetic materials using the Open Quantum Materials Database dataset.