Feature-Assisted Machine Learning for Predicting Band Gaps of Binary Semiconductors.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Computational Chemistry

Background:

• The band gap is a critical parameter for semiconductor optoelectronic devices.
• Accurate and cost-effective band gap prediction remains a challenge.
• Existing machine learning models often lack interpretability and physical grounding.

Purpose of the Study:

• To develop an interpretable and accurate machine learning model for predicting binary semiconductor band gaps.
• To integrate traditional machine learning algorithms with the sure independence screening and sparsifying operator (SISSO) approach.
• To reduce computational demands by utilizing intrinsic elemental features.

Main Methods:

• Employed a hybrid approach combining Support Vector Regression (SVR), Random Forests (RF), and Gradient Boosting Decision Trees (GBDT) with the SISSO method.
• Utilized intrinsic elemental features and Perdew-Burke-Ernzerhof (PBE) calculated band gaps.
• Applied the model to predict band gaps for 1208 theoretically stable binary compounds.

Main Results:

• Achieved enhanced accuracy and interpretability in band gap predictions for binary semiconductors.
• The model identified electronegativity as a key factor influencing material band gaps.
• Successfully predicted band gaps for a large dataset of binary compounds.

Conclusions:

• The developed hybrid ML-SISSO model offers a computationally efficient and interpretable method for band gap prediction.
• The findings deepen the understanding of physical principles governing semiconductor band gaps.
• This approach can guide the discovery and synthesis of novel semiconductor materials.