Bayesian Interatomic Potentials for Thermodynamics Stability Predictions of Icosahedral Boron Crystals

Elemental boron and icosahedral boron compounds (IBCs) are of surging interest due to their superior and versatile properties. For instance, boron suboxide (B₆O) and boron subphosphide (B₁₂P₂) derived from the α-rhombohedral boron lattice have extreme hardness and unusual electrical properties. These compounds display an extraordinary self-healing ability to repair the lattice defects generated from exposure to high-energy irradiation. Synthesis and control of the stoichiometry of IBCs are critical technical challenges. Computational tools, including Density Functional Theory (DFT) and molecular dynamics (MD) simulation, provide pathways to decipher the phase stability of IBCs. However, the structural and chemical complexity of IBCs hinder accurate DFT calculations and large-scale MD simulations.<br/>In this talk, we report a unified sparse Gaussian process (SGP), a machine learning interatomic potential, for the phase stability predictions of boron allotropes (i.e., α-B, β-B, and γ-B). To account for the variety of lattice structures, on-the-fly training was employed for precise training set generation. The trained SGP yielded good agreement with DFT on predictions of structural, thermodynamics, and vibrational properties. Insights were gained from the generated P-T diagram and free energy composition analyses. The verified SGP training strategy is also employed in the phase stability prediction of B₆O. Factors affecting the phase stability of B₆O, including defect concentration and elemental composition, were investigated using SGP interatomic potentials implemented in large-scale MD simulations. We anticipate this computational approach will aid the understanding of the phase transition of complex crystals in future studies.