Domain-informed graph neural networks: A quantum chemistry case study.

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Graph neural networks (GNNs) are powerful tools for analyzing complex systems.
• Integrating domain knowledge into GNNs can improve their accuracy and generalization capabilities.
• Estimating potential energy in chemical systems is crucial for understanding molecular and crystal behavior.

Purpose of the Study:

• To explore strategies for integrating prior domain knowledge into GNN design.
• To enhance the accuracy and generalization of GNNs for chemical system energy estimation.
• To investigate the impact of incorporating relation types and physical quantities into GNNs.

Main Methods:

• Developed GNNs that leverage knowledge of different relation types (e.g., chemical bonds) between nodes.
• Implemented specialized message production and internal state update strategies based on relation types.
• Utilized multi-task learning (MTL) to constrain learned features towards physical relevance.
• Tested the approach on three distinct GNN architectures and released new datasets.

Main Results:

• Demonstrated that integrating domain knowledge significantly improves GNN accuracy and generalization for energy prediction.
• Showcased the effectiveness of specialized message production and state update strategies.
• Validated the benefits of MTL for enhancing physical relevance of learned features.
• Confirmed the general applicability of the proposed knowledge integration methods across different GNN architectures.

Conclusions:

• Integrating domain knowledge, specifically relation types and physical quantities, is a viable strategy to enhance GNN performance in scientific applications.
• The proposed methods offer a robust framework for developing more accurate and physically meaningful GNNs for chemical system analysis.
• The public release of code and datasets facilitates further research and development in knowledge-guided GNNs.