Asynchronous Transition Across the Crystal-Melt Interface Revealed by Machine Learning Potentials

Understanding the atomic-scale structure and thermodynamics of crystal-melt interfaces is crucial in materials science. Here, we employ a DFT-based machine learning potential to investigate the crystal-melt interface, taking silicon as an example of a single-component system. We reveal that the transition is not abrupt, but spans a broad (> 12 Å) region where structural, dynamic, and thermodynamic properties evolve asynchronously. Using the local order parameter ${\overline{q}}_6$ , we identify a relative sharp structural change, sharing a common transition area with entropy/enthalpy change. The local density and self-diffusivity, however, exhibit a distinct transition area closer to the liquid side. This spatial decoupling suggests a staged melting mechanism. Furthermore, we compute the interfacial free energy via mold integration, revealing pronounced anisotropy: γ₍₁₁₁₎ < γ₍₁₁₀₎ < γ₍₁₀₀₎. This explains the prevalence of (111) facets in silicon crystal growth. The interfacial energy decreases with supercooling, lowering the nucleation barrier. Wulff construction based on our calculated energies predicts equilibrium crystal shapes consistent with experimental observations. Our work provides a comprehensive, multi-dimensional picture of the crystal-melt interface, linking atomic-scale dynamics to macroscopic crystal morphology.