Learning Atomic Multipoles: Prediction of the Electrostatic Potential with Equivariant Graph Neural Networks

Area of Science:

• Computational chemistry
• Molecular modeling
• Machine learning in science

Background:

• Classical potential-energy functions struggle with accurate electrostatic interactions.
• Fixed partial-charge approximations fail at short ranges.
• Machine-learned potentials have limitations in long-range behavior.

Purpose of the Study:

• To develop a novel method for accurate electrostatic potential description.
• To overcome limitations of existing classical and machine-learned potentials.
• To enable simultaneous accurate treatment of short- and long-range electrostatic interactions.

Main Methods:

• Introduction of an equivariant graph neural network (GNN).
• GNN predicts atomic multipoles (up to quadrupole) without quantum-mechanical calculations.
• Equivariant architecture ensures correct symmetry properties.

Main Results:

• The GNN accurately reproduces electrostatic potentials across various systems.
• The model circumvents computationally expensive quantum-mechanical computations.
• High fidelity in electrostatic potential prediction is achieved.

Conclusions:

• The proposed GNN offers a computationally efficient and accurate method for electrostatic interactions.
• Potential applications include improving machine-learned potentials and polarizable force fields.
• This approach advances the accurate modeling of electrostatic phenomena in molecules.