Data-driven learning and prediction of inorganic crystal structures.

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Solid-state physics

Background:

• Crystal structure prediction is computationally expensive.
• Quantum-mechanical methods limit the speed of traditional algorithms like ab initio random structure searching (AIRSS).
• Machine learning (ML) potentials offer a faster alternative by fitting ab initio energy landscapes.

Purpose of the Study:

• To develop Gaussian Approximation Potential-based Random Structure Searching (GAP-RSS) into a general tool for exploring configuration spaces.
• To demonstrate the capability of GAP-RSS in predicting crystalline structures.
• To explore hypothetical allotropes of phosphorus.

Main Methods:

• Developed a GAP-RSS interatomic potential model for elemental phosphorus.
• Utilized GAP-RSS to search for stable phosphorus structures without prior knowledge.
• Combined fragment analysis with GAP-RSS to explore complex structures.

Main Results:

• GAP-RSS successfully identified the orthorhombic black phosphorus (A17) structure.
• The method demonstrated the ability to "learn" known structures.
• Hypothetical 1D tubular and 3D extended phosphorus allotropes were discovered.

Conclusions:

• Machine learning potentials, like GAP-RSS, significantly accelerate crystal structure prediction.
• GAP-RSS is a versatile tool for materials discovery and design.
• ML potentials are poised to become routine computational tools in materials science.