AI agents pull 4,000 food data records from papers in 30 minutes

On a bench at Penn State, Xu Zhou measures how food materials respond to electromagnetic energy. The numbers he needs usually already exist, but they sit in thousands of papers published over decades. A team he leads has now shown that AI agents, guided by human experts, can read those papers and assemble the figures into a searchable database in a fraction of the time it used to take.

Food engineers rely on what are called physical property data, such as the dielectric properties that govern microwave heating, to build mathematical models and computer simulations. Those values are often buried in tables, figures and equations rather than in the running text that conventional data-mining tools read. The system, called AgentXtract, assigns different AI agents to different formats. They extract the numerical data, then validate, standardize and organize it, while food engineers choose the sources and make sure the data can be trusted. To choose the underlying models, the researchers compared 10 large language models from OpenAI, Google and Anthropic on accuracy, speed and cost, using data checked by hand as a reference.

The speed gap is the figure the researchers emphasize. Zhou recalls that the same work once required "an experienced graduate student or postdoctoral researcher to spend six months or more". With the agents, the team collected and organized more than 4,000 experimental data records in just 30 minutes. That speed is what made the next discovery possible. Once the data were gathered together, the team found a new and consistent relationship between the dielectric constant and the loss factor for more than 100 different foods. Zhou says it had gone unreported because collecting so much data manually was too hard. The finding comes from the team's own compilation and still needs to be checked by other researchers.

The database also serves as a knowledge base for a general-purpose AI. In a setup known as retrieval-augmented generation, the AI searches the verified entries before answering, so it can give a reference and explain the result. Zhou's example is a question about the dielectric properties of potatoes at a specific frequency, answered with the underlying data and an explanation of how those properties affect microwave heating. For food engineers, the gain is faster access to checked data, which matters for product development and process design. Zhou argues the method is not limited to food science. Its results are due in the March 2027 issue of the Journal of Food Engineering, and the article does not say when the database itself will be open to other users.