Machine-Learning Interatomic Potential for Twisted Hexagonal Boron Nitride: Polarization Analysis and Structural Insights

Recent studies on twisted two-dimensional materials have revealed intriguing ferroelectric properties, opening new doors for advanced electronic and optoelectronic applications. The twist induces structural changes due to corrugation and strain, and alters the electronic properties through atomic reconstruction. Materials modeling and understanding at this scale are crucial in the process of advancing the field, but are limited by the high computational cost of traditional methods which

calculate the electronic structure explicitly. In this context, modern machine-learning interatomic potentials, benchmarked on ab initio datasets, offer the necessary alternative. Within this framework, Gaussian approximation potentials (GAP) approach the accuracy of ab initio methods in molecular dynamics simulations, but at a far cheaper computational cost. In this study, we develop a GAP potential for twisted hexagonal boron nitride layers, enabling precise analysis of their

structural properties. After ionic relaxation and using a tight-binding model based on interatomic distances, we calculate the polarization as a function of the twist angle. Our polarization results align with previous experimental and theoretical findings.