Machine Learning Integrating Surface Features and Crystal Similarity for Exploring 2D Materials.

Area of Science:

• Materials Science
• Computational Materials Science
• Condensed Matter Physics

Background:

• Predicting properties of two-dimensional (2D) materials is challenging due to limited data and feature extraction issues.
• Existing models struggle with extrapolation and diverse material databases.

Purpose of the Study:

• To develop an advanced model for accurate 2D material property prediction.
• To enable large-scale screening and discovery of novel 2D materials.

Main Methods:

• Developed the Crystal Surface and Cluster Network (CCSN) model integrating crystal surface and cluster features.
• Utilized transfer learning from bulk material databases to enhance model performance.
• Created a prediction workflow based on crystal similarity for model training and transfer learning application.

Main Results:

• Achieved a 30% reduction in mean absolute error for bandgap prediction compared to CGCNN on a low self-similarity 2D materials database.
• Successfully predicted bandgaps for 8,218 2D crystals lacking prior data.
• Validated predictions through Density Functional Theory (DFT) calculations.

Conclusions:

• The CCSN model and associated workflow significantly improve 2D material property prediction, especially with scarce or biased data.
• This approach facilitates rapid discovery of new 2D materials.
• The methodology is generalizable to other prediction models facing data limitations.