Boron nitride (BN) is one of the most structurally stable and widely used boron compounds. However, studies on the microscopic mechanism of three-dimensional BN synthesis have so far relied solely on static analyses, lacking direct dynamic evidence. Because the synthesis is carried out under high pressure–temperature (P–T) in a sealed environment, real-time observation is feasible but often technically challenging and costly in practice. Here, we propose the use of machine-learning potential (MLP) to overcome experimental limitations and to provide unambiguous insight into material behaviors under complex and dynamic environments. In this study, we construct a first-principles-based training dataset for BN and develop an MLP within the neuroevolution potential (NEP) framework. Comprehensive stability and reliability tests confirm that the trained NEP achieves accuracy comparable to both first-principles calculations and experimental results. Using the NEP, we perform molecular-dynamics simulations starting from different stacking configurations of two-dimensional hexagonal-BN (h-BN). The results demonstrate that, under elevated P–T conditions, h-BN precursors with any stacking motif must first undergo interlayer sliding into a specific stacking sequence before compression can yield either wurtzite-BN or cubic-BN. By systematically varying simulation conditions, we construct a critical-condition diagram for BN synthesis that closely matches experimental parameters, thereby providing valuable guidance for laboratory experiments. Our work not only offers reliable theoretical insight into BN growth but also presents an effective and generalizable approach for investigating the microscopic mechanisms of chemical processes in other materials.
Wu et al. (Mon,) studied this question.