Unveiling the Phase Transformation Mechanism in BN from Large-Scale Machine Learning Molecular Dynamics

Boron nitride (BN) exhibits rich polymorphism, including cubic, hexagonal, rhombohedral, and wurtzite phases, whose dynamic transformations govern the stability and performance of BN-based materials in modern electronics, quantum photonics, and energy devices. Despite their technological importance, the atomistic mechanisms underlying its phase evolution remain poorly understood. Here, we use large-scale molecular dynamics simulations with a MACE-based machine learning interatomic potential. The potential is trained on high-fidelity density functional theory data and validated against ab initio molecular dynamics simulations to investigate phase transformation pathways in crystalline BN. We identify a layer-by-layer transformation mechanism that is initiated at heterogeneous BN interfaces, where local structural reconstruction propagates into neighboring crystalline regions. We further show that the transformation propensity depends on the interfacial structure and crystallographic orientation. These results provide atomistic insight into BN phase transformations under diverse thermodynamic conditions, while the associated MACE training workflow offers a generalizable framework for studying phase transformation mechanisms in other complex materials.