Examining Graph Neural Networks for Inorganic Crystalline Structures

Graph neural networks (GNNs) are widely used to learn the representations of crystal structures from the data. However, there lacks a systematic scheme to analyze and understand the limits of GNNs for capturing crystal structures. In this work, we propose to use human-designed descriptors as a bank of human knowledge to test whether GNNs can capture knowledge about crystal structures behind descriptors. We find that CGCNN and ALIGNN cannot learn primitive cell-level information well. Then, we suggest that, for learning extensive properties, the frequently used average pooling is not a proper choice. Finally, we propose an initial solution, hybridizing descriptors with GNNs, to improve the prediction performance of GNNs for materials properties, especially phonon internal energy and heat capacity with 90% lower errors. All the analysis can be easily extended to other deep representation learning models and human-designed descriptors. This study shows that the fields of deep representation learning and human-designed descriptors can be developed synergically.