Neural-Network-Based Molecular Dynamics Unravel Generic Multitransition States in CVD Growth of Planar Honeycomb 2D Lattices

Abstract

The chemical vapor deposition (CVD) of two-dimensional (2D) materials proceeds through a complex cascade of nonequilibrium, high-temperature surface reactions. Atomistically resolving such dynamically evolving vapor–solid interfaces remains a major challenge in both in situ experimental detection and theoretical simulations of time-dependent aggregation coupled with covalent-bond formation at the atomic scale. Here, we develop a chemistry-aware active-learning distillation (Chem-ALD) framework to construct computationally efficient, high-fidelity machine-learning force fields (MLFFs) with near quantum-chemical accuracy. By deploying the CVD-MLFF across distinct prototypical CVD systems–graphene and h-BN–on different substrates, we reveal a universal, nonclassical nucleation pathway governed by a generic multitransition states (MTSs) mechanism. Fundamentally bypassing the single-step barrier assumed by classical nucleation theory, this MTS pathway proceeds through the continuous evolution of highly fluctuating intermediates: the assembly of vapor precursors into nanochains, their progressive coalescence into branched networks, and a rate-determining transformation into critical “nano-sea-star-like” structures. In the subsequent postnucleation stage, a hexagonal nanoplatelet crystallite emerges within a fractal-like intermediate. This continuous evolution relies on the “nano-sea-star-like” topology acting as a universal kinetic funnel that connects surrounding transient states. By mapping these complex reactive landscapes at an atomistic resolution, our framework yields a unified physical picture linking localized transition-state kinetics to the macroscopic, wafer-scale synthesis of 2D single crystals.