Edge-Aware Graph Attention Model for Structural Optimization of High Entropy Carbides

TL;DR

The edge-aware graph attention model is introduced, a physics-informed graph neural network tailored for predicting relaxed atomic structures of high-entropy systems, and delivers a fast, scalable, and cost-effective alternative to DFT, enabling accelerated discovery and screening of entropy-stabilised materials.

Abstract

Predicting relaxed atomic structures of chemically complex materials remains a major computational challenge, particularly for high-entropy systems where traditional first-principles methods become prohibitively expensive. We introduce the edge-aware graph attention model, a physics-informed graph neural network tailored for predicting relaxed atomic structures of high-entropy systems. the edge-aware graph attention model employs chemically and geometrically informed descriptors that capture both atomic properties and local structural environments. To effectively capture atomic interactions, our model integrates a multi-head self-attention mechanism that adaptively weighs neighbouring atoms using both node and edge features. This edge-aware attention framework learn complex chemical and structural relationships independent of global orientation or position. We trained and evaluated the edge-aware GAT model on a dataset of carbide systems, spanning binary to high-entropy carbide compositions, and demonstrated its accuracy, convergence efficiency, and transferability. The architecture is lightweight, with a very low computational footprint, making it highly suitable for large-scale materials screening. By providing invariance to rigid-body transformations and leveraging domain-informed attention mechanisms, our model delivers a fast, scalable, and cost-effective alternative to DFT, enabling accelerated discovery and screening of entropy-stabilised materials.

Figures (6)

• Figure 1: Overall workflow of the proposed Edge-Aware GAT model from the initial structure to the final refined structure.
• Figure 2: Architecture Edge-aware GAT model. Each carbide system is represented as a crystal graph, where atoms are nodes and bonds are edges. Node features include elemental features and composition, edge attributes contain structural and chemical features. These descriptors are updated via an MLP will repeat in three blocks, and an attention mechanism selects the most influential neighbors for message passing. Final attention coefficients are used to update node features.
• Figure 3: Parity plot of predicted versus actual atomic displacement magnitudes for the test set. The close alignment of data points along the $y=x$ line indicates high agreement with DFT-relaxed positions.
• Figure 4: Probability density distribution of atomic displacement prediction errors for the test set.
• Figure 5: Mean absolute error (MAE) of predicted atomic positions along the $x$, $y$, and $z$ directions.