LiteratureAgent: Linking Perovskite Literature to Database Expansion, Knowledge Graphs, and Predictive Modeling

Scientific literature contains experimental knowledge that is difficult to combine and reuse across publications. Here, we present LiteratureAgent, a workflow that retrieves publications, extractsstructuredexperimentalinformation,retainssourcelinks,integratesrecordsintoa database, and supports knowledge-graph analysis and predictive modeling. Applied to perovskite photovoltaics, the final web-expansion campaign processed 910 publications. Across the researchgroupPDFcollectionandallweb-expansionstages,LiteratureAgentadded1,843records, expanding the Perovskite Database from 43,398 to 45,241 records. These additions came from 1,396 source identities: 1,373 identified by DOI and 23 by source filename. Adding accepted literature-derived training records improved power conversion efficiency (PCE) prediction on a separate test set of newly extracted records and produced little change on the historical test set. Stability-prediction changes were smaller and depended on the outcome being predicted. The workflow connectspublished experimentalinformation with database integration,knowledge graphs, and modeling while retaining links that allow the extracted information to be checked.