Predicting band gap from chemical composition: a comparative study of machine learning and deep learning models.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Accurate band gap prediction is vital for developing new electronic and optoelectronic materials.
• Existing methods for band gap prediction can be time-consuming and computationally intensive.

Purpose of the Study:

• To evaluate various machine learning (ML) and deep learning (DL) models for predicting inorganic semiconductor band gaps.
• To identify the most effective ML/DL approach for accurate and efficient band gap prediction.

Main Methods:

• Utilized the expt_gap dataset comprising 6,354 inorganic semiconductors.
• Applied composition-based descriptors extracted using the Magpie featurizer.
• Evaluated Linear Regression, Random Forest, XGBoost, Artificial Neural Network (ANN), and Graph Neural Network (GNN) models using fivefold cross-validation.

Main Results:

• Random Forest demonstrated the highest predictive accuracy with a Mean Absolute Error (MAE) of 0.2836 eV.
• ANN (0.3321 eV) and XGBoost (0.3575 eV) also showed strong performance.
• All tested ML/DL models significantly outperformed the Linear Regression baseline (MAE: 0.5718 eV).

Conclusions:

• Advanced ML and DL models, particularly Random Forest, effectively predict inorganic semiconductor band gaps.
• These models offer a scalable and computationally efficient alternative for accelerating materials discovery.
• Graph Neural Networks show promise for materials informatics applications.