Incorporating long-range physics in atomic-scale machine learning.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning models for atomic properties rely on locality, limiting their ability to capture nonlocal effects like long-range electrostatics.
• This limitation hinders accurate predictions for complex systems and phenomena.

Purpose of the Study:

• To develop a novel machine learning framework capable of incorporating nonlocal, nonadditive effects in atomic-scale property predictions.
• To enhance the transferability and accuracy of machine learning models by addressing the limitations of local representations.

Main Methods:

• Introduced nonlocal representations of systems, remapped as locally defined and O(3)-equivariant feature vectors.
• Developed a specific representation with asymptotic behavior matching the electrostatic potential.
• Built models for electrostatic energy, molecular dimer binding curves, and dielectric response of water.

Main Results:

• Demonstrated the framework's ability to capture nonlocal, long-range physics.
• Achieved superior performance compared to current state-of-the-art machine learning schemes.
• Successfully modeled electrostatic interactions, molecular binding, and electronic dielectric response.

Conclusions:

• The proposed method effectively integrates nonlocal physics into atomistic machine learning.
• This approach offers a significant advancement for predicting properties of materials and molecules with long-range interactions.
• The framework provides a new conceptual basis for developing more accurate and transferable machine learning models.