Multiscale machine-learning interatomic potentials for ferromagnetic and liquid iron.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Numerous interatomic potentials exist, based on analytical functions or machine learning.
• Selecting an appropriate potential requires balancing computational cost and accuracy for molecular dynamics simulations.

Purpose of the Study:

• To develop and compare four distinct interatomic potentials for iron.
• To evaluate their accuracy, transferability, and computational cost.

Main Methods:

• Developed four potentials for iron: machine-learned embedded atom method (EAM), machine-learned two- and three-body terms, machine-learned EAM with three-body terms, and Gaussian approximation potential (GAP).
• Trained all potentials on a diverse dataset of body-centred cubic and liquid iron structures computed using density functional theory (DFT).
• Evaluated potentials using cubic spline interpolation for tabulated potentials and direct implementation for GAP, spanning three orders of magnitude in computational cost.

Main Results:

• The four developed potentials exhibit a wide range of computational costs, from low to high.
• Each potential demonstrates different trade-offs between accuracy and computational efficiency.
• Performance was assessed based on transferability and the balance between predictive power and computational expense.

Conclusions:

• The study provides a comprehensive comparison of different interatomic potential types for iron.
• The findings guide the selection of appropriate potentials for molecular dynamics simulations based on specific research needs.
• Understanding these trade-offs is crucial for efficient and accurate materials simulations.