Functional machine learning modeling of electronic bandgap

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Accurate prediction of electronic bandgaps is crucial for discovering new functional materials.
• Machine learning models require robust feature sets for effective materials modeling.

Purpose of the Study:

• To investigate how functional classification of electronic bandgaps improves machine learning model performance.
• To develop and apply a systematic framework for feature selection and classification in materials science.
• To enhance the accuracy of materials property prediction for semiconductors and insulators.

Main Methods:

• Generated 518 material descriptors using the MatFeaLib Python package.
• Applied a hybrid feature selection framework to identify key descriptors.
• Utilized seven machine learning classifiers for supervised classification across different spectral regions.
• Employed hierarchical and iterative feature selection techniques for improved accuracy and pattern discovery.

Main Results:

• Achieved up to 95% accuracy in functional classification using machine learning models.
• Incorporating generalized gradient approximation gaps improved model accuracy significantly.
• SHAP analysis identified electron number, electronegativity, and bond length as key descriptors.
• Class-conditioned bandgap regression showed a 40% improvement over global regression.

Conclusions:

• Functional classification of electronic bandgaps is a powerful strategy for enhancing materials modeling.
• Distinct modeling approaches may be necessary for different material classes.
• The findings offer a new pathway for efficient functional materials discovery through improved computational methods.