Development of LLM-assisted data curation tools for the Starrydata materials science database

We developed Large Language Model (LLM)-assisted tools to accelerate curator-driven data collection from materials science publications for the Starrydata database project. We implemented two complementary tools with distinct design philosophies. The first tool is the Starrydata Auto-Suggestion for Sample Information, which generates concise English descriptions conforming to our existing database schema based on user-provided text from abstracts and experimental methods. The second tool is a schema-free dual-component system comprising the Starrydata Auto-Summary GPT and Starrydata Auto-Summary Viewer. The Auto-Suggestion tool is integrated into the Starrydata2 web platform and operates efficiently with lighter language models. The Auto-Summary GPT, demonstrated using GPT-5, processes PDF files of open-access papers and generates comprehensive JSON output capturing and summarizing all figures, tables, and experimental samples as they appear in the original papers. The companion viewer transforms this JSON data into interactive tables, enabling curators to understand complete paper structure, identify relevant data collection targets, and efficiently input figure and sample information while referencing the organized data. These tools enhance curation efficiency and represent a step toward automated scientific database construction.