GATGNN

GATGNN applies graph neural networks with augmented and global attention to predict physicochemical properties of inorganic materials.

Key Features:

• Graph Attention Mechanism: Incorporates augmented graph-attention layers (AGAT) that learn local relationships among neighboring atoms and differentiate atomic contributions.
• Global Attention Layer: Utilizes a global attention layer that assesses the overall contribution of all atoms to a material’s properties.
• PyTorch-based Implementation: Implemented as a PyTorch-based model and repository for training and inference of GNN architectures.
• Physicochemical Feature Extraction: Extracts physicochemical features from atomic/material structures to inform property prediction.
• Enhanced Predictive Performance: Demonstrated superior performance compared to state-of-the-art GNN models in predicting various material properties.

Scientific Applications:

• Materials Property Prediction: Predicts physicochemical and material properties for inorganic compounds using learned atomic interactions.
• Materials Discovery and Design: Enables analysis of atomic structure–property correlations to guide discovery and design of inorganic materials.

Methodology:

Implemented in PyTorch, GATGNN uses graph neural networks with augmented graph-attention (AGAT) layers to learn local neighbor relations and a global attention layer to aggregate contributions from all atoms for property prediction.