Graph network based deep learning of bandgaps.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning models that encode structural information enhance bandgap prediction accuracy.
• Open-access bandgap databases are crucial for expanding and updating these models.
• Graph-based deep learning methods show promise for materials property prediction.

Purpose of the Study:

• To develop a state-of-the-art multi-fidelity graph network model for predicting the bandgaps of crystalline compounds.
• To utilize a large, diverse dataset encompassing experimental and density functional theory (DFT) computed bandgaps.
• To establish a cloud-based platform for continuous model development and data accumulation.

Main Methods:

• Construction of a multi-fidelity graph network model incorporating structural information.
• Training the model on a comprehensive dataset of over 806,600 bandgap entries (experimental, low-fidelity DFT, high-fidelity DFT).
• Validation using cross-validation on high-fidelity DFT data and analysis of prediction accuracy based on crystal symmetry.

Main Results:

• The model achieved a mean absolute error of 0.23 eV for high-fidelity data in cross-validation.
• Incorporating mixed-fidelity data improved the prediction accuracy for high-fidelity bandgaps.
• Prediction error was lower for high-symmetry crystals compared to low-symmetry crystals.

Conclusions:

• The developed multi-fidelity graph network represents a significant advancement in bandgap prediction accuracy.
• The integration of diverse data fidelities and the cloud-based "Foundry" platform facilitate ongoing model improvement and data management.
• This approach enables more accurate and efficient discovery of materials with desired electronic properties.