Accelerating the application of lead-free inorganic halide perovskites in solar cells necessitates the development of novel perovskite materials with suitable bandgap widths, high stability, and environmental friendliness. This represents a crucial pathway for driving photovoltaic technology innovation and reducing reliance on conventional fossil fuels. However, traditional material development paradigms heavily depend on trial-and error experimental screening or pure density functional theory (DFT) calculations, which incur significant time and material costs.
To address these challenges, this study innovatively proposes and implements an efficient screening strategy based on the synergy between deep learning and DFT calculations. By constructing a database containing 1181 inorganic halide double perovskite materials, we systematically trained and compared the performance of five mainstream machine learning models for the bandgap prediction task: Random Forest Regression (RFR), Gradient Boosting Regression (GBR), Support Vector Regression (SVR), eXtreme Gradient Boosting Regression (XGBR), and a Deep Neural Network (DNN) model. Results demonstrate that the DNN model, leveraging its powerful nonlinear mapping capability and advantage in automatic high-dimensional feature extraction, achieved exceptional prediction accuracy on the test set, with the Mean Absolute Error (MAE) significantly reduced to 0.264 eV and the coefficient of determination (R²) reaching 0.925. Its performance was markedly superior to other compared models, highlighting the immense potential of deep learning in predicting complex material properties.
Using this optimized DNN model, this study successfully screened four promising inorganic double perovskite candidates from 55 potential materials: Cs₂GaAgCl₆, Cs₂AgIrF₆, Cs₂InAgCl₆, and Cs₂AlAgBr₆. Among them, Cs₂AgIrF₆ and Cs₂AlAgBr₆ performed particularly well, with predicted bandgaps of 1.36 eV and 1.20 eV, respectively. This range ideally matches the requirement for efficient light absorption in solar cells. Further device performance simulations revealed that the solar cell based on Cs₂AgIrF₆ achieved a simulated power conversion efficiency (PCE) of 23.71%, with an open-circuit voltage (V_OC) of 0.94 V, a short-circuit current density (J_SC) of 31.19 mA/cm², and a fill factor (FF) of 80.81%. Cs₂AlAgBr₆ also exhibited a simulated efficiency of 22.37%, corresponding to V_OC=0.78 V, J_SC=36.73 mA/cm², and FF=77.66%. Notably, both materials demonstrated high open-circuit voltages and fill factors, clearly indicating excellent carrier separation efficiency and significantly reduced nonradiative recombination losses within these materials.
In summary, this study successfully validates the significant efficacy of the deep learning-DFT synergistic strategy in accelerating the discovery of novel lead-free perovskite materials. This method not only substantially enhances the efficiency of DFT data analysis and the depth of pattern mining, overcoming some bottlenecks associated with traditional highthroughput calculations, but more importantly, it provides a practical and highly innovative approach for the rational design of high-performance, stable, and environmentally friendly lead-free perovskite solar cells, holding positive implications for advancing green, low-carbon energy technologies.