Expanding density-correlation machine learning representations for anisotropic coarse-grained particles

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Atom-centered machine learning (ML) representations are crucial for atomistic simulations.
• Current methods often assume spherical atomic interactions, limiting their use for anisotropic systems.
• Coarse-graining requires representing groups of atoms, which may not be spherical.

Purpose of the Study:

• To extend the Smooth Overlap of Atomic Positions (SOAP) ML representation for non-spherical, anisotropic particles and atom clusters.
• To introduce AniSOAP as a novel ML representation for complex systems.
• To provide insights into how molecular shape influences mesoscale behavior.

Main Methods:

• Extension of the popular Smooth Overlap of Atomic Positions (SOAP) machine learning representation.
• Development of an anisotropic SOAP (AniSOAP) descriptor.
• Application to liquid crystal systems, Gay-Berne ellipsoids, and coarse-grained benzene crystals.

Main Results:

• AniSOAP accurately characterizes liquid crystal systems.
• The method successfully predicts the energetics of anisotropic particles and coarse-grained crystals.
• Fundamental insights were derived on the influence of molecular shape on mesoscale behavior.

Conclusions:

• AniSOAP provides a powerful tool for characterizing anisotropic systems.
• The method allows for the reincorporation of atom-atom interactions often lost in coarse-graining.
• AniSOAP is proposed as a flexible, unified framework for multiscale simulations.