Learning non-local molecular interactions via equivariant local representations and charge equilibration.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Graph Neural Networks (GNNs) offer near-quantum mechanical accuracy for chemical simulations.
• Current GNNs primarily model local interactions due to their message-passing nature.
• Modeling long-range interactions like charge transfer and electrostatics remains a challenge for GNNs.

Purpose of the Study:

• To develop a novel architecture for GNN potentials capable of efficiently modeling long-range interactions.
• To generalize classical charge equilibration (Qeq) methods into a versatile GNN building block.
• To enhance the interpretability and predictive power of GNNs for complex chemical systems.

Main Methods:

• Introduction of the Charge Equilibration Layer for Long-range Interactions (CELLI) architecture.
• Generalization of classical charge equilibration principles within a GNN framework.
• Development of a model-agnostic layer compatible with equivariant GNN potentials.

Main Results:

• CELLI successfully models non-local interactions, overcoming the limitations of strictly local GNNs.
• Achieved state-of-the-art performance on benchmark systems compared to existing local models.
• Demonstrated generalization capabilities across diverse datasets and large molecular structures.

Conclusions:

• CELLI significantly extends the applicability of GNN potentials to systems requiring long-range interaction modeling.
• The explicit modeling of charges by CELLI enhances model interpretability.
• CELLI provides a computationally efficient and robust approach for advanced molecular simulations.