Perovskite databases analysis using machine learning and large language models for prediction and reverse engineering

Perovskite solar cells have reached power conversion efficiencies exceeding 26.95% in single-junction devices and 34.85% in silicon/perovskite tandems, but stability remains a major barrier to commercialization. Here, we report machine learning and large language model (LLM)-based frameworks for analyzing open-access perovskite datasets to address both efficiency and stability challenges. Classical regression models and LLMs are combined to evaluate different data representations, embedding strategies, and hyperparameter configurations for predictive and inverse design tasks. LLMs are further employed to generate fabrication procedures targeting specific key performance indicators, while clustering analysis identifies trends associated with high efficiency and stability. The proposed approach produces optimized recommendations for stack materials, additives, and processing steps across complete perovskite solar cell fabrication, achieving predicted energy yields exceeding 4,000 kWh/m 2 . These results demonstrate the potential of LLM-assisted frameworks for accelerating the discovery and optimization of perovskite solar cells.