Machine learned potential for defected single layer hexagonal boron nitride.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine learned interatomic potentials (MLIPs) are crucial for simulating materials at laboratory-relevant scales.
• Simulating hexagonal boron nitride (h-BN) with defects and grain boundaries requires accurate interatomic potentials.

Purpose of the Study:

• To develop a robust MLIP for simulating temperature-dependent properties of single-layer h-BN, including defects and grain boundaries.
• To validate the MLIP's accuracy against ab initio calculations and experimental data.

Main Methods:

• Utilized a local equivariant deep neural network (Allegro) to create the MLIP.
• Trained the MLIP on a dataset of ~30,000 h-BN structures with defects, generated using density functional theory (DFT) and ab initio molecular dynamics at various temperatures.
• Validated the MLIP by comparing predicted energies, forces, phonon dispersion, and vibrational density of states with DFT results.

Main Results:

• The MLIP achieved low prediction errors for potential energies (4 meV/atom) and forces (60 meV/Å).
• It accurately reproduced phonon dispersion curves and vibrational density of states for pristine h-BN.
• Simulations of a 4|8 grain boundary revealed distinct activation barriers for initial motion (~2.2 eV) versus subsequent motion (~0.42 eV), and predicted a scaled mobility at 1500 K in reasonable agreement with experimental values.

Conclusions:

• The developed MLIP is robust and suitable for reliable simulations of defect structures and dynamics in single-layer h-BN.
• The MLIP's ability to predict grain boundary mobility demonstrates its potential for advancing materials simulations.
• This work facilitates the study of complex phenomena in 2D materials like h-BN.