Hybrid Graph-Machine Learning Framework for Accurate and Interpretable Band Gap Prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Accurate electronic band gap prediction is crucial for discovering new semiconducting and energy materials.
• Conventional density functional theory (DFT) methods are computationally expensive and lack scalability.
• There is a need for efficient and accurate predictive models in materials informatics.

Purpose of the Study:

• To develop a hybrid artificial intelligence (AI) framework for high-accuracy, interpretable, and computationally efficient band gap prediction.
• To integrate graph-based deep learning embeddings with classical machine learning algorithms.
• To leverage physically meaningful crystal descriptors for enhanced model performance.

Main Methods:

• A hybrid AI framework combining CGCNN, MEGNet, and SchNet embeddings with crystal descriptors (electronegativity, crystal system, space group, spin-orbit coupling).
• Training using optimized gradient-boosting and neural network architectures on 136,000 crystal structures from the Materials Project database.
• Utilizing SHAP for interpretability analysis to understand feature importance.

Main Results:

• The hybrid model achieved R² = 0.921, MAE = 0.191, and MSE = 0.155.
• Outperformed conventional DFT, classical models, and standalone graph neural networks.
• Demonstrated comparable accuracy to state-of-the-art ALIGNN with lower computational cost and better generalization.
• Interpretability analysis revealed metallicity and magnetic site features as key predictors.

Conclusions:

• The synergy between deep structural embeddings and classical algorithms offers a powerful and scalable approach for materials informatics.
• The proposed framework provides a foundation for multiproperty prediction, transfer learning, and inverse materials design.
• This interpretable AI approach accelerates the discovery and design of novel materials.