Generalized representative structures for atomistic systems.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Existing methods for generating representative atomic structures are often limited to crystalline systems and fixed lattices.
• Reproducing essential characteristics like chemical disorder in large alloys within small structures is challenging.
• A more general description of atomic systems requires comprehensive atomic environment descriptors.

Purpose of the Study:

• To present a novel data-driven method for generating small atomic structures that represent arbitrary material systems, phases, or ensembles.
• To overcome the limitations of previous methods by not restricting structures to fixed lattices.
• To enable efficient and systematic generation of representative atomic structures with high chemical and spatial complexity.

Main Methods:

• Utilizing the atomic cluster expansion (ACE) basis for a complete set of atomic environment descriptors.
• Implementing the Generalized Representative Structure (GRS) method to generate structures based on ACE descriptor distributions.
• Employing optimization algorithms for efficient generation of diverse material representations.

Main Results:

• Demonstrated the ability to generate systematically improvable representations for crystalline systems, amorphous materials, and liquids.
• Highlighted reduced representations of atomistic machine-learning training datasets with equivalent information content.
• Showcased small (40-72 atom) representations of liquid phases and the potential for novel structure generation.

Conclusions:

• The GRS method provides a powerful, generalizable approach for creating representative atomic structures across various material types.
• This method offers significant advantages over existing data-driven and high-symmetry restricted techniques.
• GRS facilitates efficient materials modeling, data compression, and the exploration of novel material structures.