A crystal graph to vector approach for predicting magnetic properties.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting material properties computationally is crucial for materials discovery.
• Current deep graph network models excel on large datasets but struggle with limited data, particularly for magnetic properties.
• Existing methods often lack interpretability and scalability.

Purpose of the Study:

• Introduce CG-Vec, a crystal graph-to-vector framework.
• Develop a machine learning approach that is accurate, interpretable, and efficient, especially in data-scarce scenarios.
• Enable reliable prediction of diverse material properties, including magnetic characteristics.

Main Methods:

• Replaced iterative message passing in graph networks with compact, interpretable descriptors.
• Coupled these descriptors to conventional machine learning algorithms.
• Validated the framework on diverse datasets for magnetic, formation energy, and band gap properties.

Main Results:

• CG-Vec matches deep graph network accuracy on large datasets.
• CG-Vec substantially outperforms deep graph networks in data-scarce regimes.
• Achieved reliable predictions for magnetization (ferromagnetic and ferrimagnetic) and Curie temperature.
• Demonstrated competitive performance for formation energy and band gap prediction.

Conclusions:

• Vectorized representations offer a practical and scalable alternative to deep architectures for materials modeling.
• CG-Vec enables efficient and interpretable prediction of complex material properties.
• The framework shows broad applicability across different material property types.