Machine-Learning Interatomic Potential for Twisted Hexagonal Boron Nitride: Accurate Structural Relaxation and Emergent Polarization

Abstract

The emerging ferroelectric properties of two-dimensional (2D) heterostructures are at the forefront of science and prospective technology. In moir\'e bilayers, twisting or heterostructuring causes local atomic reconstruction, which even at picometer scale, can lead to pronounced ferroelectric polarization. Accurately determining this reconstruction utilizing ab initio methods is unfeasible for the relevant system sizes, but modern machine-learning interatomic potentials offer a viable solution. Here, we present the Gaussian Approximation Potential for twisted hexagonal boron nitride (hBN) layers validated against ab initio datasets. This approach enables the precise analysis of their structural properties, which is particularly relevant at small twist angles. We couple the structural information to a tight-binding model based on accurate interatomic positioning, and determine the twist-dependent polarization, yielding results that closely align with previous experimental findings - even at room temperature. This methodology enables further studies that are unattainable otherwise and is transferable to other 2D materials of interest.