Graph Transformer Model Integrating Physical Features for Projected Electronic Density of States Prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Traditional methods for predicting projected density of states (PDOS) often use graph convolutional networks (GCN) and Graph Attention Networks (GAT).
• These models have limitations in accurately capturing complex electronic structures.

Purpose of the Study:

• To evaluate the performance of the Graph Transformer (GT) model for PDOS prediction.
• To investigate the impact of incorporating physics-informed features, such as valence electron counts, on PDOS prediction accuracy.

Main Methods:

• Utilized PDOS data from the Materials Project.
• Compared the predictive accuracy of Graph Transformer (GT) against Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN).
• Enhanced the GT model by including valence electron counts (s, p, d, f orbitals) as input features.

Main Results:

• The Graph Transformer (GT) model consistently outperformed both GAT and GCN models in PDOS prediction accuracy.
• The inclusion of valence electron counts as a feature significantly improved the predictive performance of the GT model.
• The study demonstrated a novel framework for enhancing PDOS prediction.

Conclusions:

• The Graph Transformer (GT) model represents a significant advancement in predicting projected density of states (PDOS).
• Physics-informed feature engineering, specifically using valence electron counts, is crucial for improving the accuracy of deep learning models for electronic property prediction.
• This approach offers a pathway for more accurate predictions of PDOS and other electronic properties in materials.