From Papers to Standardized Metadata: AI-supported Workflow for Extracting Knowledge on Agricultural Longterm Experiments

Agricultural long-term experiments (LTEs) provide valuable multi decadal data for crop modeling and global change research, yet their broader reuse remains limited by the narrow focus of many studies and incomplete adherence to FAIR principles. The LTE Overview Map (https://bonares.de/ltfe/) has compiled metadata for about 680 LTEs worldwide, supporting the reuse of both published and unpublished data and facilitating synthesis studies that combine multiple LTEs. However, the metadata are still incomplete and require substantial manual curation. Here, we present TrialHarvest-AI, an automated pipeline that combines large language models (LLMs) with rigorously defined Pydantic data schemas that guide and control the AI in extracting structured information directly from scientific literature. Implemented as a Python workflow, the system parses papers associated with each LTE, extracts relevant experimental descriptors, and translates them into standardized, interoperable metadata records. In addition, it identifies the research context and specific objectives of each publication using LTE data – information that provides initial indicators for data fitness for purpose and helps track LTE data reuse. Quality control begins with structural validation, where Pydantic ensures that all extracted fields follow the required format. This is followed by semantic checks that test whether the information makes logical and domain specific sense. Finally, we compare results of 3-5 LLMs and use an AI as judge process that rates the accuracy and reliability of each extraction, supporting transparent and trustworthy metadata generation. This AI supported approach greatly accelerates metadata curation, improves data discoverability, and supports a more systematic understanding of how LTE data are used across the agricultural research community.