Machine learning-driven prediction of perovskite properties for PV applications

This article presents a study of the fundamental properties of β-CsYX 3 (Y = Ca, Cd, Ge, Pb, Sn, Sr; X = I, Br, Cl) perovskite materials using machine learning methods. Six algorithms, such as linear regression, Ridge regression, robust regression, k-nearest neighbors regression, ElasticNet, and Lasso regression, are used to predict the bandgap, Debye temperature, and unit cell volume. Based on MAE, MSE, and R 2 metrics, it was determined that linear regression yields the best results for a number of the properties under investigation. However, analysis of the RMSE revealed that, for certain properties, lower RMSE values were achieved using robust regression and Ridge regression. Thus, model performance depends on both the target property and the chosen evaluation metric. Residual analysis also demonstrated relatively stable prediction behaviour across all models considered. Overall, the results confirm the potential of machine learning methods for the accelerated screening and discovery of perovskite materials with favourable electronic and structural characteristics. The predicted values for band gap, Debye temperature, and unit cell volume serve as useful descriptors for the preliminary assessment and comparison of candidate materials for optoelectronic applications.