Spin-informed universal graph neural networks for simulating magnetic ordering.

Area of Science:

• Materials Science
• Computational Physics
• Machine Learning

Background:

• Discovering magnetic materials is computationally expensive using density-functional theory (DFT).
• Current universal machine-learning interatomic potentials (uMLIPs) lack magnetic ordering prediction capabilities.
• Existing methods struggle with the high computational cost of determining ground-state magnetic configurations.

Purpose of the Study:

• To develop a machine learning framework for efficient prediction of magnetic ordering.
• To extend the capabilities of universal machine-learning interatomic potentials (uMLIPs) to include magnetic properties.
• To accelerate the screening and discovery of novel magnetic materials.

Main Methods:

• A data-efficient, spin-informed graph neural network framework was developed.
• The framework incorporates spin degrees of freedom and preserves physical symmetries.
• A closed-loop anomaly detection approach was implemented for dataset enhancement.

Main Results:

• The framework significantly speeds up density-functional theory (DFT) calculations by providing better initial guesses for magnetic moments.
• Accurate determination of ground-state magnetic ordering in bulk materials was achieved.
• The model demonstrated generalization capabilities for predicting magnetic ordering in surfaces.

Conclusions:

• The developed framework enhances the utility of uMLIPs for magnetic materials research.
• This approach accelerates the computational discovery of magnetic materials.
• Anomaly detection improves the accuracy and robustness of machine learning models in materials science.