2026-10-16 ペンシルベニア州立大学（Penn State）
＜関連情報＞
- https://www.psu.edu/news/research/story/can-ai-help-discover-new-food-physics-researchers-pair-ai-human-expertise
- https://www.sciencedirect.com/science/article/abs/pii/S0260877426003651
食品特性データのためのエージェント型AI：文献からデータベース、そして新たな発見へ Agentic AI for food property data: From literature to database to new discovery
Xu Zhou, Yi Zhang, Shyam S. Sablani, Shaojin Wang, Juming Tang
Journal of Food Engineering Available online: 12 September 2026
DOI:https://doi.org/10.1016/j.jfoodeng.2026.113327
Highlights
- Multi-agent LLM system converts food literature into structured databases.
- Ten LLMs reached >95% precision and >99% value accuracy on dielectric data.
- Database showed a new relationship between dielectric constant and loss factor.
- RAG uses database data to provide scientific, quantitative answers.
Abstract
Food property data are important for product development and process design and, more recently, for AI model training, but most data are scattered in literature. Building food property databases manually is time-consuming and difficult to scale. We developed AgentXtract, an agentic AI workflow that automatically transforms scientific papers into databases. Using dielectric properties as a case study, three specialized large language model (LLM) agents collaborate in a coordinated workflow, with different roles in paper screening and interpretation, data extraction, and record standardization and assembly. Ten LLMs from OpenAI, Google, and Anthropic were benchmarked against an expert-curated reference dataset. All models achieved precision above 95% and value accuracy above 99%, but recall, API costs, and processing times varied. The best-performing model was then selected to construct the full database. AgentXtract generated 4276 records for 142 food materials in approximately 30 min with an API cost of $9. Analysis of the compiled database revealed new relationships between dielectric constant and loss factor that have not been reported previously. The resulting database also served as the knowledge source for a prototype retrieval-augmented generation (RAG) AI model. Users can ask questions about dielectric properties and microwave heating in plain language, and RAG searches the database to find relevant data and provides answers with references. This study demonstrates that AI can efficiently and accurately transform fragmented food literature into structured databases and also support data-driven discovery in food engineering.
