Predicting photovoltaic efficiency of two-dimensional Janus materials for applications in solar energy harvesting: A combined first-principles and machine learning study

Abstract

The present paper reports a combined first-principles density functional theory (DFT) and machine learning (ML) study to predict the photovoltaic efficiency of two-dimensional Janus materials. Initially, the electronic and optical properties of four Janus systems Sc₂CFBr, Sc₂CFCl, Sc₂CHCl and Sc₂CHF are investigated using DFT. All four compounds exhibit indirect semiconducting behaviour with HSE06 band gaps ranging from 1.61 to 1.88 eV and strong visible light absorption. The associated average charge density, effective mass of charge carriers, and exciton binding energy have also been estimated. The photovoltaic performance parameters including short-circuit current density, open-circuit voltage, fill factor, and maximum power density highlight maximum photovoltaic efficiency (${\eta }_{\max }$) for Sc₂CFBr (23.83 %) followed by Sc₂CFCl (20.22 %), Sc₂CHCl (19.41 %), and Sc₂CHF (17.96 %) compounds. Extending the analysis, a dataset of 562 Janus materials is screened and refined to 343 candidates with optimal photovoltaic band gaps through confusion matrix analysis. Correlation between key feature descriptors and target variable ${\eta }_{\max }$ have been extracted using Pearson correlation heatmap. Ten ML algorithms, including five deep learning and five shallow learning models, are employed. The hybrid model integrating Convolutional Neural Network and Kernel Ridge Regressor exhibits superior predictive performance of ${\eta }_{\max }$ with R² values 0.96 (training) and 0.95 (testing). This study demonstrates the efficacy of ML, especially deep-shallow unified model, in predicting ${\eta }_{\max }$ of Janus materials which may bear promising applications towards next-generation high-efficiency and sustainable photovoltaics.