Crystal nucleation and growth dynamics of aluminum via quantum-accurate MD simulations

Abstract

The experimental study of crystal nucleation and growth in deeply supercooled liquids is challenging because of the minuscule size of the critical nuclei and the short timescales involved. Computational simulations have become powerful tools to overcome these challenges; however, they are often biased by predefined interatomic potential functions, which may lack transferability, oversimplify complex atomic interactions, and struggle to accurately capture phase transitions. In this work, we present a comprehensive molecular dynamics (MD) study of crystal nucleation and growth in aluminum, using a recently developed machine learning (ML) model trained exclusively on liquid-phase DFT configurations – without any prior knowledge of solid properties and structures. This ML model accurately reproduces key thermodynamic and structural properties of real aluminum, including heat capacity, lattice parameter, and melting point. We investigate spontaneous and seeded crystallization in the temperature ranges T=500–540 K and T=600–790 K, identifying emergent crystalline clusters using the pair entropy fingerprint method, independent of predefined crystal patterns. The homogeneous nucleation rate, J, was calculated by Classical Nucleation Theory (CNT) using MD-derived properties, without any fitting parameters. There was an excellent agreement between theoretical predictions and direct MD-derived values of J, corroborating the validity of CNT. Additionally, the computed solid-liquid interfacial free energy was consistent with experimental estimates. Furthermore, crystal growth dynamics from both spontaneously formed nuclei and inserted seeds were accurately described by the Turnbull-Fisher (TF) model, using simulation-derived parameters. This finding was corroborated by an additional analysis of crystal growth in a two-million-atom Lennard-Jones (LJ) liquid at four temperatures. The macroscopic growth rates predicted by the TF model showed good consistency with independently computed values for flat LJ surfaces. Notably, at 10% supercooling, the theoretical growth rate derived from MD data on nanosized LJ nuclei aligns exceptionally well with recent experimental measurements for pure argon (a proxy for an LJ system). Thus, this research bridges the gap between experiments, theory, and simulations in crystal nucleation and growth. Overall, this study demonstrates that an ML-driven, crystal-unbiased model can accurately capture the kinetics and thermodynamics of crystallization, validating two classical phenomenological theories at the atomic scale.