Atomistic Line Graph Neural Network for Improved Materials Property Predictions (2106.01829v3)

Abstract: Graph neural networks (GNN) have been shown to provide substantial performance improvements for atomistic material representation and modeling compared with descriptor-based machine learning models. While most existing GNN models for atomistic predictions are based on atomic distance information, they do not explicitly incorporate bond angles, which are critical for distinguishing many atomic structures. Furthermore, many material properties are known to be sensitive to slight changes in bond angles. We present an Atomistic Line Graph Neural Network (ALIGNN), a GNN architecture that performs message passing on both the interatomic bond graph and its line graph corresponding to bond angles. We demonstrate that angle information can be explicitly and efficiently included, leading to improved performance on multiple atomistic prediction tasks. We ALIGNN models for predicting 52 solid-state and molecular properties available in the JARVIS-DFT, Materials project, and QM9 databases. ALIGNN can outperform some previously reported GNN models on atomistic prediction tasks by up to 85% in accuracy with better or comparable model training speed.

• The paper presents ALIGNN, a new GNN architecture that integrates bond distances and angles through a line graph framework.
• ALIGNN achieves up to a 43.6% reduction in formation energy prediction MAE compared to benchmark models across multiple datasets.
• The study’s comprehensive evaluation across electronic and structural properties establishes a robust pathway for advancing computational materials design.

Atomistic Line Graph Neural Network for Enhanced Materials Property Prediction

The paper introduces the Atomistic Line Graph Neural Network (ALIGNN), an innovative Graph Neural Network (GNN) architecture tailored for improved material property predictions. ALIGNN aims to address a gap present in existing GNN models used for material science, which often emphasize atomistic distances while underestimating the importance of bond angles. The enhancement of model accuracy achieved by integrating bond angle information into the GNN structure is the focal point of this paper.

Key Contributions

1. Novel GNN Architecture: ALIGNN effectively incorporates both bond distances and bond angles in its predictive modeling framework. This is achieved through an inventive use of a line graph, whereby edges of the original atomic graph (representing bonds) are treated as nodes, and bond angles are introduced as edges in this secondary graph. The alternated message-passing between the conventional bond graph and this line graph allows ALIGNN to capture atomistic structural nuances better than traditional approaches.
2. Superior Performance: The ALIGNN model demonstrates significant performance gains across various datasets, including the JARVIS-DFT, Materials Project, and QM9 databases. Specifically, it achieves superior mean absolute error metrics (MAE) for properties such as formation energies and band gaps, outperforming existing models like Crystal Graph Convolutional Neural Network (CGCNN) and MatErials Graph Network (MEGNet). For instance, ALIGNN reduces the MAE of formation energy predictions by up to 43.6% compared to CGCNN.
3. Extensive Evaluation: The utility of ALIGNN is comprehensively validated through systematic evaluation on central tasks concerning molecular and solid-state properties such as electronic band structures, formation energies, and piezoelectric coefficients. The paper meticulously analyzes these properties across varied datasets, underscoring the robustness and versatility of the ALIGNN architecture.

Practical and Theoretical Implications

The explicit incorporation of bond angles through line graphs marks a paradigm shift in atomistic predictions, facilitating a higher fidelity to real-world material structures. This approach provides a theoretical foundation and practical pathway for enhancing the predictive capabilities of GNNs in material science. The improvements in prediction accuracy could profoundly impact computational material design, aiding in the discovery of new materials with optimized electronic, magnetic, and structural properties.

Future Prospects

This research opens myriad avenues for future exploration within the field of AI-driven materials science. A significant future direction could involve scaling this architecture to accommodate even more intricate material systems, potentially synergizing with quantum computing simulations for prediction enhancement. Additionally, further diversification of datasets and integration with real-world experimental results could bolster the applicability and accuracy of GNN predictions in various domain-specific material engineering tasks.

In conclusion, ALIGNN represents a pivotal advancement in the use of graphical and machine learning algorithms for materials property prediction, offering enhanced accuracy and efficiency through the novel inclusion of angular information in atomistic modeling.