Accelerating materials property prediction via a hybrid Transformer Graph framework that leverages four body...

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Machine learning accelerates inorganic materials property prediction.
• Challenges include data scarcity for specific properties and thermodynamic stability assessment.

Purpose of the Study:

• To develop a robust framework for predicting energy-related and mechanical properties of inorganic materials.
• To address data scarcity in mechanical property prediction using transfer learning.
• To improve the accuracy and interpretability of materials property prediction models.

Main Methods:

• Utilized a Graph Neural Network with composition-based and crystal structure-based architectures.
• Incorporated four-body interactions to capture periodicity and structural characteristics.
• Employed a transfer learning scheme to handle data scarcity, particularly for mechanical properties.

Main Results:

• Achieved accurate predictions for energy-related properties (total energy, energy above convex hull, band gap) and mechanical properties (bulk and shear modulus).
• Outperformed state-of-the-art models in 8 materials property regression tasks.
• Demonstrated superior prediction of local atomic environments and global structural features.

Conclusions:

• The proposed framework effectively predicts diverse material properties, including those with limited data.
• Transfer learning and interpretable model components enhance materials design and discovery.
• The approach offers a significant advancement in efficient and accurate materials property prediction.