Representing local atomic environment using descriptors based on local correlations.

Area of Science:

• Materials Science
• Computational Chemistry
• Statistical Learning

Background:

• Statistical learning is crucial for understanding material properties and identifying new materials like catalysts and electronic materials.
• Efficiently analyzing large datasets requires effective descriptors to screen atomic environments and learn complex energy landscapes.
• Discovering descriptors for material properties and their functional dependencies is challenging and time-consuming.

Purpose of the Study:

• To develop a novel framework for generating robust descriptors of local atomic environments.
• To enable efficient and accurate learning of material properties from large datasets.
• To overcome the limitations of existing methods in capturing intricate geometric features.

Main Methods:

• Generated descriptors based on many-body correlations (two-body, three-body, four-body, and higher).
• Descriptors capture intrinsic geometric features of local atomic environments.
• Descriptors are invariant to global translation, rotation, reflection, and atomic index permutations.
• Evaluated descriptors using overlap integrals.

Main Results:

• The developed descriptors effectively capture the intrinsic geometric features of local atomic environments.
• Demonstrated the ability to successfully reconstruct complex structures containing 10-25 atoms.
• Achieved a significant improvement over previous methods in structure reconstruction.

Conclusions:

• The many-body correlation-based descriptors provide a powerful tool for statistical learning in materials science.
• This framework enhances the ability to predict and design materials with desired properties.
• The method offers a more efficient and accurate approach to analyzing complex atomic structures.