Machine learning interatomic potentials for aluminium: application to solidification phenomena.

Area of Science:

• Materials Science
• Computational Physics
• Chemical Engineering

Background:

• Accurate modeling of solidification requires interatomic potentials that capture both solid and liquid states.
• Ab initio molecular dynamics (AIMD) is limited by computational cost for large-scale, long-time simulations of nucleation and relaxation.

Purpose of the Study:

• To develop a classical molecular dynamics (MD) potential capable of simulating solidification phenomena at larger scales than AIMD.
• To investigate homogeneous nucleation mechanisms in elemental aluminum under various conditions.

Main Methods:

• A high-dimensional neural network potential was trained using AIMD-generated configurations relevant to solidification.
• The dataset included diverse crystalline and liquid states of aluminum across various temperatures and pressures.
• Classical MD simulations were performed using the trained potential on systems up to one million atoms.

Main Results:

• The neural network potential accurately reproduced structural, dynamic, and thermodynamic properties of liquid and undercooled aluminum.
• Simulations revealed homogeneous nucleation mechanisms in both face-centered cubic (fcc) and body-centered cubic (bcc) phases.
• A single-step nucleation process was observed for both fcc and bcc phases.

Conclusions:

• The developed neural network potential enables efficient and accurate large-scale simulations of solidification.
• The study provides unprecedented insights into homogeneous nucleation mechanisms in elemental aluminum.