Defect graph neural networks for materials discovery in high-temperature clean-energy applications

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Predicting defect formation enthalpies is crucial for materials discovery but computationally expensive.
• Traditional methods require creating and relaxing defected atomic structures, limiting high-throughput screening.

Purpose of the Study:

• To develop a fully automated graph neural network (GNN) approach for predicting defect formation enthalpies.
• To enable rapid screening of materials for energy applications by replacing computationally intensive density functional theory (DFT) calculations.

Main Methods:

• Trained a defect graph neural network (dGNN) model using DFT reference data for vacancy defects in oxides.
• Integrated the dGNN model with thermodynamic calculations for free energy assessments.
• Applied the dGNN model to screen oxides in the Materials Project database.

Main Results:

• The dGNN model accurately predicts defect formation enthalpies from ideal crystal structures, eliminating the need for explicit defect modeling.
• Successfully connected zero-Kelvin defect enthalpies to high-temperature process conditions for solar thermochemical hydrogen production.
• Demonstrated the model's applicability to arbitrary structures and potential generalizability to other defect types.

Conclusions:

• The dGNN approach offers a computationally efficient and automated method for materials discovery.
• This technique accelerates the identification of novel materials for clean energy applications.
• The dGNN framework is adaptable and can be extended to various defect types and advanced architectures.