Discovery of germanium-based hybrid perovskites with high photovoltaic performance with machine-learning.

The discovery of stable, high-performance lead-free hybrid perovskite materials remains a major challenge for next-generation photovoltaic technologies. Here, a high-throughput computational framework integrating machine learning, deep learning, and density functional theory (DFT) calculations was adopted to accelerate the discovery and validation of lead-free hybrid perovskites. An optimized XGBoost model was firstly employed to screen approximately 180k charge-neutral perovskite candidates according to formation-energy and bandgap criteria. The resulting 48 kinds of Ge-based candidates were subsequently validated using CGCNN and UPET models, leading to the identification of (CH)4NH2GeGeCl6, C3H4NSGeGeBr6, C5NH12GeGeBr6 and C6OH13GeGeBr6 as promising photovoltaic materials, with HSE + SOC-calculated bandgaps of 1.5137, 1.7515, 1.6778 and 1.1680 eV, respectively. These results demonstrate that integrating data-driven prediction with first-principles validation provides an effective strategy for accelerating the discovery of next-generation photovoltaic materials.