Cartesian message passing neural networks for directional properties: Fast and transferable atomic multipoles.

Area of Science:

• Computational chemistry
• Machine learning for materials science

Background:

• Message Passing Neural Networks (MPNNs) are effective for atomic property modeling.
• Traditional MPNNs struggle with directional properties like Cartesian tensors.
• Atom-centered multipoles are vital for accurate ab initio force fields.

Purpose of the Study:

• To develop a modified MPNN (CMPNN) capable of predicting atom-centered multipoles.
• To assess the CMPNN's performance on a large dataset of chemical structures.
• To evaluate the impact of predicted multipoles on electrostatic energy calculations.

Main Methods:

• Implementation of a modified Cartesian Message Passing Neural Network (CMPNN).
• Training and validation using a dataset of 46,623 chemical structures with high-quality atomic multipoles.
• Analysis of prediction accuracy for atomic charges, dipoles, and quadrupoles.

Main Results:

• The CMPNN accurately predicts atom-centered charges, dipoles, and quadrupoles.
• Errors in predicted multipoles minimally impact multipole-multipole electrostatic energies.
• The model demonstrates capability in capturing conformational dependencies of electronic structure.

Conclusions:

• The CMPNN is a viable tool for predicting atom-centered multipoles.
• This method enhances the accuracy of ab initio force fields.
• Enables on-the-fly calculation of conformation-dependent atomic multipoles during simulations.