Graph Comparison of Molecular Crystals in Band Gap Prediction Using Neural Networks

Area of Science:

• Materials Informatics
• Computational Materials Science
• Solid-State Chemistry

Background:

• Graph representation is crucial for material structure analysis in material informatics.
• Molecular and crystal structures of molecular crystals can be represented as graphs.
• A comparative analysis of these graph representations for predictive modeling is lacking.

Purpose of the Study:

• To compare the effectiveness of molecular graphs versus crystal graphs for predicting material properties.
• To evaluate which graph representation yields higher prediction accuracy for the band gap of molecular crystals.

Main Methods:

• Utilized graph-based machine learning models to represent molecular crystals.
• Compared prediction accuracy for band gaps using both molecular graph and crystal graph representations.
• Performed quantitative error analysis to assess the contribution of crystal structure information.

Main Results:

• Crystal graph representations demonstrated superior prediction accuracy for band gaps compared to molecular graph representations.
• Error analysis indicated crystal graphs reduced prediction errors by approximately 0.4 times compared to molecular graphs.
• A moderate correlation was observed between crystal structure features and band gap prediction accuracy.

Conclusions:

• Crystal graph representations are more effective than molecular graph representations for band gap prediction in molecular crystals.
• Incorporating crystal structure information significantly enhances the accuracy of material property predictions.
• This study provides a quantitative basis for selecting optimal graph representations in material informatics.