Simple and Efficient Equivariant Message-Passing Neural Network Model for Non-local Potential Energy Surfaces

Area of Science:

• Computational materials science
• Machine learning in physics
• Quantum chemistry

Background:

• Machine learning potentials (MLPs) excel in atomistic simulations.
• Describing nonlocal interactions beyond local environments remains a challenge for MLPs.

Purpose of the Study:

• To propose a simple and efficient equivariant model, EquiREANN, for representing nonlocal potential energy surfaces.
• To address the challenge of efficiently modeling nonlocal interactions in atomistic simulations.

Main Methods:

• Developed EquiREANN, an equivariant model based on a message-passing framework.
• Utilized linear combinations of atomic orbitals as fundamental descriptors.
• Iteratively updated invariant orbital coefficients and equivariant orbital functions.

Main Results:

• EquiREANN accurately describes potential energy variations due to nonlocal structural changes.
• The model achieves high accuracy with computational costs comparable to invariant models.
• Demonstrated the model's ability to capture subtle nonlocal effects.

Conclusions:

• EquiREANN offers an effective solution for representing nonlocal potential energy surfaces in atomistic simulations.
• The proposed approach provides a generalized method for adapting equivariant message-passing to other descriptors.
• This work advances the capability of machine learning in modeling complex material behaviors.