Haris Suhendar
Accurate prediction of electronic properties such as band gaps is crucial for accelerating the discovery of functional perovskite materials. In this study, we employed a dataset of 45,570 perovskite compounds, with structural and electronic properties derived from density functional theory (DFT) calculations. The dataset exhibits a highly skewed distribution of band gap values, dominated by metallic or nearly metallic compounds with zero band gaps, posing a challenge for machine learning (ML) modeling. To address this, we benchmarked four ML algorithms XGBoost, Random Forest (RF), LightGBM (LGBM), and Support Vector Regression (SVR) for band gap prediction. Performance was evaluated using mean absolute error (MAE), root mean squared error (RMSE), and the coefficient of determination (R2) across training and testing sets. Ensemble-based methods (XGBoost, RF, LGBM) consistently outperformed SVR, with LGBM achieving the best testing performance (MAE = 0.413 eV, RMSE = 0.478 eV, R2 = 0.791), closely followed by XGBoost (MAE = 0.429 eV, RMSE = 0.491 eV, R2 = 0.784). Correlation plots confirmed the superior predictive accuracy of ensemble models, while SVR showed substantial deviations and systematic underestimation of higher band gaps. Further analysis of XGBoost revealed that predictive performance saturates after incorporating 1/420 features, with testing R2 stabilizing around 0.75-0.80, indicating that a compact set of descriptors captures most of the predictive information. These findings highlight the effectiveness of gradient boosting methods, particularly XGBoost and LGBM, for robust band gap prediction in complex perovskite systems, thereby providing a reliable framework to accelerate materials discovery. © Published under licence by IOP Publishing Ltd.
Department of Physics, Universitas Negeri Jakarta, Indonesia
Research at a Glance
Register to unlock
Topics & SDG Alignment
Register to unlock
Collaboration
Register to unlock
Author Profile (Selected)
Register to unlock
References Overview
Register to unlock
Journal & Source
Register to unlock
Metadata & Integrity
Register to unlock