X2-GNN: A Physical Message Passing Neural Network with Natural Generalization Ability to Large and Complex Molecules.

Area of Science:

• Computational chemistry
• Machine learning
• Materials science

Background:

• Neural networks excel at molecular property prediction but struggle with generalization to larger molecules.
• Increased molecular size leads to greater structural diversity and complex interactions, challenging existing models.

Purpose of the Study:

• To develop an E(3) invariant graph neural network (GNN) named X2-GNN for enhanced molecular property prediction.
• To improve the generalization capabilities of neural networks from small to large molecules by incorporating physical insights.

Main Methods:

• Introduced X2-GNN, an E(3) invariant message passing GNN.
• Integrated atomic orbital overlap integrals and core Hamiltonians to provide physical insights.
• Employed an attention mechanism to enhance learning efficiency.

Main Results:

• X2-GNN demonstrated effective generalization to larger molecules (tens of heavy atoms) when trained on smaller datasets (QM9).
• Achieved credible per-atom error rates in molecular property predictions.
• Showcased high accuracy in potential energy surface modeling and bond dissociation energy prediction within subseconds.

Conclusions:

• X2-GNN exhibits scalability and broad applicability for molecular property prediction.
• Integrating data-driven methods with electronic structure theory knowledge is crucial for advancing computational chemistry.