Machine Learning for the Prediction of Chalcohalide Energy Band Gaps: A Tool for Identifying New Photovoltaic and Photosensitive Materials

Chalcohalide compounds have been recognized as outstanding photoabsorbers attractive for use in photocatalysis, solar cells, second harmonic generation, and photodetectors. The reliable experimental band gaps (Eg) of chalcohalides were collected manually from original research papers. The features for machine learning (ML) were generated using information on the chemical composition of chalcohalide material and fundamental properties of chemical elements. The minimum, maximum, and weighted mean of these parameters were used as train data. Different regression models of traditional machine learning and deep learning were trained and cross-validated. The hyperparameter tuning was performed to gain the ML performance. The quality of regression was evaluated by calculating the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). The best results (RMSE = 0.47 eV, MAE = 0.34 eV, R2 = 0.71) were obtained for the Extreme Gradient Boosting Regression (XGBR) algorithm. Additional validation was accomplished by predicting the energy band gap values of eight new chalcohalides recently reported in the literature within the last few years. The average relative error of Eg prediction for these compounds attained a low value of 16% in the case of the most efficient XGBR model. The developed ML method can be applied as an alternative way to complex and time-consuming Density Functional Theory (DFT) calculations or expensive experimental measurements.