Descriptor engineering in machine learning regression of electronic structure properties for 2D materials.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting electronic properties of 2D materials is crucial for materials discovery.
• Existing methods may be computationally intensive or lack accuracy for novel materials.

Purpose of the Study:

• To develop novel material descriptors for predicting the band gap and work function of 2D materials.
• To enhance the accuracy and efficiency of machine learning models for 2D material property prediction.

Main Methods:

• Constructed new material descriptors by vectorizing property matrices and using empirical property functions.
• Combined these novel features with database-based features for tree-based machine learning models.
• Utilized extreme gradient boosting for predicting band gap and work function.

Main Results:

• Achieved high prediction accuracy with R-squared values > 0.9 and Mean Absolute Errors (MAE) < 0.23 eV.
• Extreme gradient boosting yielded R-squared of 0.95 (band gap) and 0.98 (work function), with MAEs of 0.16 eV and 0.10 eV, respectively.
• Demonstrated improved performance over database features alone and reduced overfitting.

Conclusions:

• The developed descriptor-based method efficiently predicts electronic properties of 2D materials.
• Hybrid features and ensemble models provide a robust framework for materials property prediction.
• Offers a guideline for engineering descriptors for efficient 2D material property prediction.