Homogeneous nucleation of undercooled Al–Ni melts via a machine-learned interaction potential

Johannes Sandberg, Thomas Voigtmann, Émilie Devijver, N. JakseThe Journal of Chemical Physics 2025 · DOI · arXiv

This research paper, recently published by a team from Germany and France, leverages machine learning to bridge a long-standing gap in materials science: how to simulate the birth of a crystal from a molten metal with both the speed of classical physics and the accuracy of quantum mechanics.

TL;DR

Researchers developed a high-dimensional neural network potential (HDNNP) to simulate the solidification of Al-Ni alloys. Their large-scale simulations reveal that both pure Nickel and equiatomic AlNi nucleate in a single step, correcting previous misconceptions from simpler models and showing that the "seeds" of crystals are far more irregular than classical theories suggest.

What Problem Does This Paper Address?

Understanding how a liquid turns into a solid—a process called homogeneous nucleation—is fundamental to controlling the microstructure (and thus the strength and durability) of industrial materials. However, studying this at the atomic level is notoriously difficult.

There are two traditional ways to simulate this, both with major drawbacks:

  1. Ab initio simulations: These use first-principles quantum mechanics. They are highly accurate but computationally "expensive," meaning they can only simulate a few hundred atoms for a few picoseconds—nowhere near enough to see a rare nucleation event.
  2. Classical empirical potentials: These use simplified "ball-and-spring" physics. They are fast enough for large scales but often lack the accuracy to describe complex chemical interactions, sometimes leading to incorrect predictions about which crystal structures will form.

The authors tackle this by using Machine Learning Interatomic Potentials (MLIPs) to get the best of both worlds: quantum-level accuracy at a fraction of the computational cost.

Methods

The team built a High-Dimensional Neural Network Potential (HDNNP) based on the Behler-Parrinello framework.

  • Training: They used Ab Initio Molecular Dynamics (AIMD) to generate a "gold standard" dataset of how atoms behave at various temperatures and compositions.
  • The "Brain": The neural network was trained to recognize atomic environments (using "symmetry functions" as descriptors) and predict the potential energy of the system.
  • Large-Scale MD: Once trained, the potential was plugged into the LAMMPS molecular dynamics engine to simulate 135,000 atoms—a scale large enough to observe the spontaneous formation of crystal nuclei.
  • Analysis: They used Common Neighbor Analysis (CNA) and bond-order parameters to identify the specific crystal structures (FCC, BCC, HCP) as they emerged from the melt.

Datasets

The study utilized several data sources for training and validation:

  • Primary Training Data: New AIMD trajectories for Al-Ni at five different concentrations (0%, 25%, 50%, 75%, and 100% Nickel).
  • Structural Benchmarks: Crystalline phases (stable and unstable) sourced from the Materials Project database to ensure the model recognized solid structures as well as liquids.
  • Experimental Validation: The model's predictions were compared against real-world neutron diffraction data (structure factors), self-diffusion coefficients, and shear viscosity measurements from existing literature.

Key Insights and Ideas

  • The "Single-Step" Revelation: Previous classical models (like EAM) suggested that pure Nickel nucleates in two steps—first forming a temporary BCC structure before settling into its final FCC form. This ML model shows that Ni actually nucleates in a single step directly into FCC. This highlights how sensitive our understanding of physics is to the accuracy of the underlying "force field" used in simulations.
  • Chemical Affinity as a Driver: In the 50/50 AlNi mix, the strong chemical attraction between Al and Ni atoms in the liquid state pre-orders the melt. This leads to a direct, single-step nucleation into the B2 (ordered BCC) structure, rather than the FCC structures seen in the pure elements.
  • Challenging Classical Theory: Classical Nucleation Theory (CNT) often assumes that the initial "seeds" of a crystal are perfect spheres. The simulations showed that these nuclei are actually irregular and non-spherical, which suggests that standard mathematical models for nucleation rates may need to be refined.

Key Results

  • Accuracy: The ML potential showed "nearly perfect match" with ab initio data for structural properties and transport coefficients (diffusion and viscosity).
  • Nucleation Timing:
    • AlNi (50/50): Nucleation began around 170 ps after deep undercooling (at 1300K).
  • Pure Ni: Nucleation was much slower, beginning around 1.77 ns (at 1200K).
  • Critical Size: The authors estimated the "critical nucleus size" (the size at which a crystal seed becomes stable enough to grow rather than dissolve) for AlNi to be between 80 and 120 atoms.

Limitations and Future Work

  • Force Fitting: The current model was trained only on energy levels, not atomic forces. While this maintained consistency with the authors' previous work on pure Aluminum, they acknowledge that including force data in the training could further improve accuracy.
  • Compositional Scope: While the potential is accurate across the whole Al-Ni range, the authors only performed large-scale nucleation simulations for two specific points (pure Ni and 50/50 AlNi).
  • Future Extensions: The authors suggest that their "embedded feature selection" method could be applied next to optimize the neural network for even more complex, multi-component alloy systems.

Summary updated

Abstract

Homogeneous nucleation processes are important for understanding solidification and the resulting microstructure of materials. Simulating this process requires accurately describing the interactions between atoms, which is further complicated by chemical order through cross-species interactions. The large scales needed to observe rare nucleation events are far beyond the capabilities of ab initio simulations. Machine learning is used for overcoming these limitations in terms of both accuracy and speed, by building a high-dimensional neural network potential for binary Al-Ni alloys, which serve as a model system relevant to many industrial applications. The potential is validated against experimental diffusion, viscosity, scattering data, and its melting temperature. It is applied to large-scale molecular dynamics simulations of homogeneous nucleation, specifically for Al-Ni at equiatomic composition and for pure Ni. Pure Ni nucleates in a single step into an fcc crystal phase, in contrast to previous results obtained with a classical empirical potential. This highlights the sensitivity of nucleation pathways to the underlying atomic interactions. Our findings suggest that the nucleation pathway for AlNi proceeds in a single step toward a B2 structure, which is discussed in relation to the pure elements counterparts.
BibTeX
@article{sandberg2025homogeneous,
  author = {Johannes Sandberg and Thomas Voigtmann and Émilie Devijver and N. Jakse},
  title = {{Homogeneous nucleation of undercooled Al–Ni melts via a machine-learned interaction potential}},
  journal = {The Journal of Chemical Physics},
  year = {2025},
  doi = {10.1063/5.0299431},
}
Figure