Band Gap Prediction in 2D Materials via Machine Learning and Explainable AI: A Physics Inspired Approach

Two dimensional (2D) materials have become the leaders in the next generation semiconducting applications due to their outstanding electronic and optical characteristics. Predicting the electronic band gap which is a crucial property influencing device performance is computationally expensive if determined through traditional density functional theory (DFT) method. In this study, we present a machine learning (ML) framework that leverages rich features derived from the JARVIS-DFT database to predict DFT-computed band gaps for ~1000 monolayer 2D materials. Also, a high-quality physics-inspired feature engineering methodology and explainable AI (XAI) models are used to improve the accuracy of predictions and their interpretability. Then, we compare 12 kernel-based regression models on a pipeline of cross-validation (5-fold) and grid search optimization. ExtraTrees regressor is the best performing model achieving R2of around 0.80 and RMSE approximately 0.48 eV on a hold-out test set. In order to make scientific transparency, we use SHAP analysis to identify the main structural and electronic features behind predictions. The combination of informatics of materials and XAI provides a scalable and interpretable alternative to DFT that speeds up the discovery of materials in the 2D material design space.