Can AI help discover new food physics? Researchers pair AI with human expertise

UNIVERSITY PARK, Pa. — Information about food properties that determine quality, safety and sensory characteristics is critical for food product development and processing design, but it is spread across thousands of scientific papers written over decades, making it difficult for food engineers to access and analyze data. Now, a team led by researchers at Penn State have shown how artificial intelligence (AI) agents, guided by human expertise, can extract and organize experimental data from scientific literature into a searchable database that helps reveal previously unreported physical relationships. The same database also serves as a knowledge base for general-purpose AI, helping it provide more accurate, evidence-based answers to scientific questions.

The team’s findings are available online ahead of publication in the March 2027 issue of the Journal of Food Engineering.

The idea for the research occurred to study corresponding author Xu Zhou, assistant professor in Penn State’s Department of Food Science, during his interview for a faculty position at Penn State in late 2025. When he learned about work done in the College of Agricultural Sciences to connect artificial intelligence with Penn State Extension in developing the search assistant Tilva, he realized a similar concept could solve a problem he had encountered.

“I use AI a lot, and I can see it is getting more and more powerful, but sometimes, even when AI can search scientific papers, it still cannot give me the answers I need for food science questions,” he said. “I realized one reason is that the foundational model may not have enough high-quality food data in its training, or when it searches online, it does not always know which data are good. To determine that, it needs not only scientific knowledge but also to be provided with hands-on experimental experience.”

Food engineers need reliable physical property data for mathematical models and computer simulations, but those data are scattered through decades of publications and can take a great deal of time to find and collect, Zhou explained.

“In this work, we explored how AI agents and human experts can work together, with food engineers providing the expertise needed to identify scientific data sources and ensure data quality and reliability, and AI handling large-scale data collection,” he said.

In the study, the researchers created AgentXtract, in which AI agents work together to interpret scientific papers. Unlike traditional text-based data mining, much of the valuable experimental data in food engineering is reported in tables, figures and mathematical equations. AI agents with different skills work together to extract numerical data from different formats, then validate, standardize and organize the information into a database.

To test AgentXtract’s performance, the researchers evaluated 10 different large language models from OpenAI, Google and Anthropic to power the AI agents. They compared the models’ performance — accuracy, speed and cost — to interpret scientific papers, understand experimental data and extract relevant numerical information, using manually verified data from human experts as a reference.

“In the past, I would need an experienced graduate student or postdoctoral researcher to spend six months or more,” Zhou said. “Now, with AI agents, we collected and organized more than 4,000 experimental data records in just 30 minutes.”

The research also led to an unexpected discovery, Zhou noted. By analyzing the collected data, the researchers found a previously unreported mathematical relationship between two dielectric properties, which describe how food interacts with electromagnetic energy, such as microwaves.

“What excited me more was that when we put all data together, we observed a new and consistent relationship between dielectric constant and loss factor for more than 100 different foods,” Zhou said. “This relationship was not reported before, because it was so hard to collect and organize so much data manually. Now, AI makes this possible.”

Finally, in aretrieval-augmented generation — or RAG — application, the researchers connected the database to a question-answering AI to evaluate its performance.

“Someone could ask, for example, ‘What are the dielectric properties of potatoes at a particular frequency?’” Zhou said. “And the system quickly retrieves relevant database entries and provides an answer with references and explains how those dielectric properties affect microwave heating. Today’s large language models are already very powerful, but one way to make them even more useful is to provide them with high-quality scientific data verified by human experts. Our work shows how we can do that, not just for food science, but also for many other areas of research.”

Contributing to the research were Yi Zhang, Frederik Sr. and Faith E. Rasmussen Career Development Professor in Food Science at Penn State; Shyam Sablani and Shaojin Wang, Department of Biological Systems Engineering, Washington State University; and Juming Tang, Department of Industrial & Systems Engineering, University of Washington.

This research was supported by start-up funding provided to Xu Zhou by Penn State and Anthropic’s AI for Science Program. This content is the sole responsibility of the authors and does not necessarily reflect the views of the funders.