Neural-network interatomic potential for grain boundary structures and their energetics in silicon

T Yokoi and Y Noda and A Nakamura and K Matsunaga, PHYSICAL REVIEW MATERIALS, 4, 014605 (2020).

DOI: 10.1103/PhysRevMaterials.4.014605

Artificial neural-network (ANN) interatomic potentials for simulating atomic structures and energetics of grain boundaries (GBs) in silicon were constructed and integrated into structural optimization and molecular dynamics (MD) algorithms. A training dataset including various atomic environments of symmetric tilt GBs was generated by performing density-functional-theory (DFT) calculations. The ANN potential after training was found to be capable of approximating the potential-energy surface at GBs even with dangling bonds and large atomic displacements at high temperatures, which cannot be well reproduced with empirical interatomic potentials. Additionally, reliability of the ANN potential for molecular simulations was evaluated. GB structures optimized or equilibrated by the ANN molecular simulations were also energetically lower for DFT calculations, without significant errors. The ANN potential is therefore expected to greatly reduce structural- optimization iterations and required time steps to acquire stable or equilibrium GB structures in MD simulations, enabling us to address even large-scale systems of general GBs in silicon, with high accuracy and low computational cost.

Return to Publications page