Node-equivariant message passing for efficient and accurate machine learning interatomic potentials.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Computational Biophysics

Background:

• Equivariant message-passing (MP) models offer high fidelity for first-principles data in materials science, biophysics, and catalysis.
• Current equivariant MP models face computational and memory limitations due to expensive tensor operations, hindering large-scale simulations.

Purpose of the Study:

• To develop a novel node-equivariant message-passing (NEMP) framework.
• To reduce the computational and memory footprint of equivariant MP models.
• To enable large-scale and long-time simulations with high accuracy.

Main Methods:

• Proposed a novel NEMP framework performing equivariant operations between a central node and a virtual summed node.
• The virtual node encodes structural information of neighboring nodes.
• Evaluated NEMP across diverse systems: molecules, extended systems, and universal potential benchmarks.

Main Results:

• NEMP achieves comparable or superior accuracy to existing edge-equivariant MP models.
• Demonstrated a 1-2 order of magnitude reduction in memory and computational costs.
• NEMP exhibits computational efficiency comparable to local descriptor-based models.

Conclusions:

• The NEMP framework significantly enhances the efficiency of machine learned interatomic potentials.
• NEMP overcomes the limitations of previous models, enabling previously inaccessible large-scale simulations.
• This work paves the way for more extensive computational studies in various scientific domains.