Band-Gap Regression with Architecture-Optimized Message-Passing Neural Networks.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Graph-based neural networks, particularly message-passing neural networks (MPNNs), show promise for predicting solid-state properties.
• Accurate prediction of material properties is crucial for discovering new materials.

Purpose of the Study:

• To train an MPNN for classifying materials as metallic or semiconducting/insulating using density functional theory (DFT) data.
• To optimize MPNN architecture and hyperparameters for predicting band gaps of nonmetallic materials.
• To evaluate ensemble methods and uncertainty quantification for improved predictive accuracy.

Main Methods:

• Utilized density functional theory (DFT) data from the AFLOW database.
• Trained a message-passing neural network (MPNN) for material classification.
• Performed neural-architecture search to optimize MPNNs for band gap prediction.
• Ensembled top-performing models and employed Monte Carlo dropout for uncertainty quantification.

Main Results:

• The MPNN successfully classified materials as metallic or semiconducting/insulating.
• An ensemble of MPNNs significantly outperformed the best single model in predicting band gaps.
• Ensembling proved superior to Monte Carlo dropout for uncertainty quantification.

Conclusions:

• MPNNs are effective tools for predicting electronic properties of solids.
• Ensemble modeling enhances prediction accuracy and reliability for materials properties.
• The study analyzed the model's domain of applicability across various material characteristics.