Composition and Structure Based GGA Bandgap Prediction Using Machine Learning Approach
This study focuses on developing precise machine learning (ML) regression models for predicting energy bandgap values based on chemical compositions and crystal structures. The primary aim is to match the accuracy of predictions derived from GGA-PBE calculations and validate them through density functional theory (DFT)-based band structure calculations. We assessed eight standalone ML regression models, including AdaBoost, Bagging, CatBoost, LGBM, RF, DT, GB, and XGB. These models were analyzed for their ability to predict GGA-PBE bandgap values across diverse material structures and compositions, using a dataset containing bandgap values for 106,113 compounds. Additionally, we constructed four ensemble models using the stacking method and seven using the bagging method. These ensemble models incorporated RidgeCV and LassoCV to explore if ensemble techniques could enhance prediction accuracy. The dataset was divided into subsets of varying sizes: 10,000, 25,000, 50,000, and 100,000 entries. We determined feature importance through permutation techniques and established a correlation coefficient matrix using the Pearson correlation method. The Random Forest (RF) model emerged as the top performer among standalone models, achieving an R2 value of 0.943 and an RMSE value of 0.504 eV. Bagging regression demonstrated improved performance across different dataset sizes with streamlined feature selection. Ensemble models, particularly bagging, consistently outperformed standalone models, achieving the best R2 value of 0.948 and an RMSE value of 0.479 eV in the test dataset. Using the best-performing model, we predicted bandgap values for new half-Heusler compounds with 18 valence electron counts. These predictions were successfully validated using accurate DFT calculations. DFT calculations indicated that the newly predicted compounds are narrow bandgap semiconductors with dynamic stability.