Formula Graph Self-Attention Network for Representation-Domain Independent Materials Discovery

Area of Science:

• Materials Science
• Machine Learning
• Computational Chemistry

Background:

• Machine learning (ML) for materials property prediction relies heavily on effective material representations.
• Current ML models are often limited to either structure-based or stoichiometry-only descriptors, hindering broader applicability.
• Graph neural networks (GNNs) show promise but face challenges due to the distinct nature of existing descriptors.

Purpose of the Study:

• To introduce a unified 'formula graph' concept bridging stoichiometry-only and structure-based material descriptors.
• To develop a self-attention integrated GNN capable of processing these unified formula graphs.
• To enhance material embeddings for improved transferability and prediction accuracy in materials science.

Main Methods:

• Development of the novel 'formula graph' representation unifying stoichiometric and structural information.
• Implementation of a self-attention mechanism within a GNN architecture to process formula graphs.
• Evaluation of the model's performance, sample efficiency, and convergence compared to existing methods.

Main Results:

• The proposed GNN with formula graphs generates transferable material embeddings across descriptor domains.
• The model demonstrates superior performance, sample efficiency, and faster convergence over prior structure-agnostic and structure-based models.
• Successful application in predicting the complex dielectric function and identifying potential epsilon-near-zero materials.

Conclusions:

• The formula graph concept effectively unifies diverse material representations for ML.
• The self-attention integrated GNN offers a powerful and versatile tool for materials property prediction.
• This approach advances the discovery of new materials with targeted functionalities, such as epsilon-near-zero phenomena.