Physics-Inspired Equivariant Descriptors of Nonbonded Interactions

Area of Science:

• Computational materials science
• Machine learning for chemistry
• Atomistic modeling

Background:

• Machine learning (ML) methods for atomistic modeling often rely on locality, neglecting crucial long-range (LR) interactions like electrostatics and dispersion.
• This neglect limits the accuracy of ML models for certain material and molecular properties.

Purpose of the Study:

• To extend the long distance equivariant (LODE) framework to consistently incorporate diverse long-range interactions.
• To develop a coherent approach for data-driven modeling of matter that handles arbitrary nonbonded interactions.

Main Methods:

• Developed new atom-centered features for the LODE framework to represent long-range interactions.
• Provided a physical interpretation of these features using multipole expansion for efficient implementation.
• Applied the generalized LODE framework to toy systems and molecular dimers.

Main Results:

• Demonstrated the framework's ability to handle diverse LR interactions in a consistent manner.
• Achieved simpler and more efficient implementations through multipole expansion interpretation.
• Successfully applied the method to various molecular systems, validating its performance.

Conclusions:

• The generalized LODE framework offers a coherent approach to treat arbitrary two- and many-body nonbonded interactions in ML-based matter modeling.
• This extension significantly enhances the capability of machine learning models to capture complex interactions in materials and molecules.