Accuracy and limitations of machine-learned interatomic potentials for magnetic systems: A case study on Fe-Cr-C

Area of Science:

• Computational Materials Science
• Condensed Matter Physics
• Materials Informatics

Background:

• Machine-learned interatomic potentials (MLIPs) are increasingly vital for atomistic simulations.
• Applying MLIPs to magnetic materials is challenging due to the need to account for spin fluctuations.
• The Fe-Cr-C system is technologically significant, making it a key test case for magnetic alloy simulations.

Purpose of the Study:

• To investigate the impact of training data (nonmagnetic vs. spin-polarized) on MLIP accuracy for magnetic alloys.
• To develop efficient strategies for constructing MLIPs for magnetic materials, including transfer learning.
• To benchmark the performance of specialized MLIPs against state-of-the-art foundation models.

Main Methods:

• Constructed two deep machine learning potentials (DP-NM, DP-M) using the DeePMD framework, trained on nonmagnetic and spin-polarized DFT data, respectively.
• Employed a transfer-learning strategy: pretraining on nonmagnetic data followed by fine-tuning on spin-polarized data.
• Validated MLIPs against experimental data and benchmarked against foundation models (MACE, GRACE, DPA3) for dynamic and static properties.

Main Results:

• The nonmagnetic-trained potential (DP-NM) accurately reproduced dynamic properties like viscosity and melting points.
• The spin-polarized-trained potential (DP-M) excelled at describing static properties like density, particularly for Fe-rich alloys.
• Transfer learning significantly reduced computational costs for spin-polarized data by over an order of magnitude.
• Foundation models, while accurate after fine-tuning, were slower for molecular dynamics simulations.

Conclusions:

• Nonmagnetic training data is adequate for dynamic properties of paramagnetic melts.
• Spin-polarized training is essential for accurate static properties of ferromagnetic phases.
• Clear design principles for MLIPs in magnetic alloys are established, guiding efficient development and clarifying the role of foundation models and transfer learning.