Atomic structure of boron resolved using machine learning and global sampling.

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• The atomic structure of beta-boron (β-B) remains uncertain after decades of study, hindering understanding of its exotic photoelectric properties.
• The vast configurational space of self-doped boron crystals makes determining the true atomic configuration extremely difficult with current methods.

Purpose of the Study:

• To explore the potential energy surface of β-B and identify governing rules for interstitial site filling.
• To develop an accurate and efficient computational model for predicting boron crystal structures.

Main Methods:

• Combined machine learning (ML) with stochastic surface walking (SSW) global optimization to explore β-B's potential energy surface.
• Developed a novel neural network (NN) potential using advanced structural descriptors to accurately model complex boron bonding environments.

Main Results:

• Identified 15,293 distinct configurations from over 200,000 visited minima, revealing key rules for interstitial site occupancy.
• Found only 40 low-energy configurations within 7 meV/atom of the global minimum, with many discovered for the first time.
• Classified low-energy structures into three skeleton types and six doping patterns, showing a preference for specific interstitial sites, notably the B19 site.

Conclusions:

• The B19 interstitial site significantly influences β-B properties, with its occupancy becoming dominant at higher temperatures due to large vibrational entropy.
• The novel SSW-NN architecture is a powerful tool for solving complex material phenomena and accelerating materials genome database development.