Multi-scale approach for the prediction of atomic scale properties

Area of Science:

• Condensed matter physics
• Quantum mechanics
• Computational chemistry

Background:

• Machine learning models for quantum mechanics often rely on local atomic environments, limiting their ability to capture long-range physical effects.
• Existing methods struggle with electrostatic interactions, quantum delocalization, and other phenomena that extend beyond short-range atomic contributions.

Purpose of the Study:

• To develop a novel multi-scale scheme that integrates both local and non-local information within a single framework.
• To overcome the inherent limitations of purely local approaches in modeling complex physical phenomena.

Main Methods:

• A data-driven approach was used to construct a multi-scale scheme combining local and non-local features.
• The simplest form of these features was shown to correspond to a multipole expansion of permanent electrostatics.
• The model's capability to handle delocalized and collective effects was explored.

Main Results:

• The multi-scale scheme successfully models interactions driven by electrostatics, polarization, and dispersion.
• Demonstrated ability to capture the cooperative behavior of dielectric response functions.
• Validated across diverse applications including molecular physics, surface science, and biophysics.

Conclusions:

• The developed multi-scale approach effectively bridges the gap between local and non-local interactions in condensed matter.
• This framework offers a powerful tool for accurate prediction of quantum mechanical observables, enhancing the capabilities of machine learning in materials science and beyond.