PerovskiteNet: A text-mined database of perovskite solar cells with fabrication details using large language models and data-driven performance prediction

Perovskite solar cells (PSCs) have advanced rapidly, but the experimental literature remains difficult to use at scale because device stacks, fabrication details, and performance metrics are reported in heterogeneous prose and supplements. We introduce PerovskiteNet, a large text-mined PSC database curated with a language-model-based multi-agent system. PerovskiteNet aggregates 10,258 papers from 2021--2025, records the best-performing device in each paper, and represents every entry with 97 fields covering reference metadata, cell definition, device stack, synthesis, fabrication, J-V metrics, and stability. The curation pipeline is paired with a STEAD benchmark using independent judge agents, yielding 89.71% accuracy and 91.79% F1 for the selected annotator. We use the database to summarize recent architectural and compositional trends and, in a controlled SnO$_2$/Spiro case study, train predictive models on 2,445 devices with cross-validated $R^2>0.5$. These results show how evaluated language-model curation can turn fast-moving scientific literature into machine-readable infrastructure for materials discovery.