Ensemble Learning and Graph Neural Networks for High Throughput Screening of Non-Toxic, Thermally Stable Hybrid Perovskites for Solar Energy

This study introduces an artificial intelligence framework for accelerating the discovery of stable, lead-free hybrid organicinorganic double perovskites for solar energy applications.We combined a pre-trained Atomistic Line Graph Neural Network (ALIGNN) with gradient boosting ensembles to predict three critical properties: formation energy, bandgap, and Debye temperature.The ALIGNN model was trained on 8,000 crystal structures and achieved mean absolute errors of 0.011 eV per atom for formation energy, 0.094 eV for bandgap, and 10.5 K for Debye temperature.The gradient boosting models provided complementary accuracy and interpretability, particularly for bandgap classification.Using this pipeline, we screened 8,412 candidate compounds and identified KAgBiBr as a promising material with a bandgap of 1.34 eV, a Debye temperature of 402 K, and a formation energy of -2.31 eV per atom.These values suggest long-term thermal stability and high photovoltaic potential without toxic lead.Compared with density functional theory calculations, our approach reduces computational cost by more than 90 percent while maintaining predictive fidelity.The framework offers a scalable path toward rapid identification of practical solar absorber materials and could significantly shorten the timeline for developing safe and efficient perovskite photovoltaics.