AI Agents and Experts Reveal Novel Food Physics Relationship

Penn State University researchers have developed a human-guided artificial intelligence workflow that can mine decades of food science literature, turning scattered experimental tables, figures and equations into a searchable database capable of revealing new physical relationships and improving AI answers about food properties. Led by Xu Zhou, assistant professor of food science, the team created AgentXtract, a system in which multiple AI agents collaborate to locate, validate, standardize and organize numerical data. Food experts guide the process by identifying trustworthy sources and confirming data reliability, addressing a major weakness of general-purpose AI models when handling specialized scientific evidence. To evaluate the approach, the researchers tested 10 large language models from OpenAI, Google and Anthropic for accuracy, speed and cost in interpreting papers and extracting dielectric property data related to microwave and electromagnetic interactions. Zhou said the workflow collected and organized more than 4,000 experimental records in 30 minutes, work that previously could require an experienced graduate student or postdoctoral researcher for six months or longer. The consolidated database also uncovered a previously unreported mathematical relationship between the dielectric constant and loss factor across more than 100 foods, a pattern too broad to detect through manual data assembly. The team further integrated the verified database into a retrieval-augmented generation question-answering system, allowing users to ask for properties such as the dielectric behavior of potatoes at specific frequencies and receive referenced, evidence-based explanations relevant to food processing and safety. The findings are available online ahead of publication in the March 2027 issue of the Journal of Food Engineering.