Machine learning-aided discovery of bismuth-based transition metal oxide double perovskites for solar cell applications

Due to their importance in semiconductor device designing, especially in photovoltaic solar cells and light emitting diodes, methods that can promptly and reliably forecast material's bandgap (E-g) and its character, viz., direct or indirect, are in demand. In this context, data-driven machine learning (ML) methodologies are considered promising. In the present work, several ML models were developed using easy-to-find instrumental variables such as unit-cell volume, structural parameters (a, b, c, alpha, beta, gamma), space group, number of constituent atoms, and standard atomic properties (viz., atomic number, atomic mass, ionisation energy, electronegativity) to forecast the bandgap and its character for double perovskites. The LGBMRegressor and XGBClassifier models were identified to best predict the magnitude and nature of the bandgap with an accuracy of similar to 0.89 and 0.95, respectively. Subsequently, the above models were employed to predict the bandgap for novel bismuth-based transition metal oxide double perovskites. The accuracy of the present models, especially over the range of 1.2-1.8 eV, makes them particularly suitable for designing bismuth-based double perovskites for photovoltaic applications.

Automated summary

In this study, machine learning models were developed to predict the bandgap and its character of double perovskite materials, with LGBMRegressor and XGBClassifier models identified as the best predictors. These models were further employed to predict the bandgap of novel bismuth-based transition metal oxide double perovskites, showing high accuracy, especially in the range of 1.2-1.8 eV.