Band Gap Prediction of Two-Dimensional Materials Using a Gradient-Boosted Feature Selection Approach

Area of Science:

• Materials Science
• Computational Materials Science
• Condensed Matter Physics

Background:

• Two-dimensional (2D) inorganic crystals are crucial for advanced electronic and optoelectronic devices due to unique properties like high carrier mobility and tunable band gaps.
• Data-driven approaches are increasingly used for predicting material properties, offering computationally efficient alternatives to traditional methods, especially for screening candidate materials.
• Existing data-driven models often focus on 3D bulk materials, leaving a gap in predictive capabilities for 2D layered, van der Waals, and ultrathin film materials.

Purpose of the Study:

• To develop and validate data-driven models for predicting key properties of 2D inorganic materials: thermodynamic stability, metallicity, and electronic band gap.
• To ensure the developed models are fully interpretable, providing insights into feature relevance for property prediction.
• To assess the models' performance on both in-distribution and out-of-distribution datasets for robust material screening.

Main Methods:

• Utilized open-source databases (Alexandria_2D, C2DB, MC2D, 2DMatpedia) to compile a dataset of 2D materials.
• Employed chemically relevant elemental, physical, and compositional features as input for the models.
• Applied statistical and gradient-boosted feature selection to reduce dimensionality and enhance model interpretability using SHapley Additive exPlanations (SHAP).

Main Results:

• Classifiers for thermodynamic stability and metallicity achieved high accuracy (89.7%) and F1-scores (0.832 and 0.870, respectively).
• The band gap regressor demonstrated strong performance on the in-distribution test set with an R-squared of 0.883.
• The band gap model showed moderate predictive ability on an out-of-distribution dataset, with an R-squared of 0.334, highlighting the challenges and potential for further improvement in predicting unseen material properties.

Conclusions:

• Data-driven models, coupled with effective feature selection, provide accurate and interpretable predictions for 2D material properties.
• These explainable surrogate models are effective for high-throughput screening, accelerating the discovery of novel 2D materials for electronic and optoelectronic applications.
• The study validates the utility of machine learning in materials science for predicting thermodynamic stability, metallicity, and electronic band gaps in 2D materials.