DenseGNN: universal and scalable deeper graph neural networks for high-performance property prediction in crystals and molecules

Abstract

Abstract Modern deep learning-driven generative models have made it possible to design millions of hypothetical materials. However, to sift through these candidate materials and identify promising new materials, we need fast and accurate models for predicting material properties. Graph neural networks (GNNs) have emerged as a current research hotspot due to their ability to directly operate on the graph representations of molecules and materials, enabling comprehensively capturing key information and exhibiting outstanding performance in predicting material properties. Nevertheless, GNNs still face several key problems in practical applications: firstly, existing nested graph network strategies, while able to incorporate critical structural information such as bond angles, significantly increase the number of trainable parameters in the model, leading to a substantial rise in training costs; secondly, extending GNN models to broader fields such as molecules, crystalline materials, and catalysis, as well as adapting to small datasets, remains a challenge; finally, the scalability of GNN models are limited by the over-smoothing problem. To address these problems, we propose the DenseGNN model, which combines dense connectivity network (DCN), hierarchical node-edge-graph residual networks (HSN), and Local structure Order Parameters Embedding (LOPE) strategies, aiming to create a universal, scalable and efficient GNN model. We have achieved state-of-the-art (SOAT) performance on multiple datasets including JARVIS-DFT, Materials Project, QM9, Lipop, FreeSolv, ESOL, and OC22, demonstrating the generality and scalability of our approach. By fusing DCN and LOPE strategies into GNN models in the fields of computer, crystal materials, and molecules, we have significantly enhanced the performance of models such as GIN, Schnet, and Hamnet on material datasets like Matbench. The LOPE strategy optimizes the embedding representation of atoms, enabling our model to train efficiently at a minimal level of edge connections, significantly reducing computational costs, shortening the time required to train large GNNs, while maintaining accuracy. Our technique not only supports the construction of deeper GNNs, avoiding performance degradation problems seen in other models, but is also applicable to a wide range of applications requiring large deep learning models. Furthermore, our study demonstrates that by utilizing structural embeddings from pre-trained models, our model not only outperforms other GNNs in crystal structure distinguishment, but also approaches the standard X-ray diffraction (XRD) method.