The Pulse
Penn State’s AgentXtract Links Dielectric Properties Across Foods
Penn State researchers used AgentXtract to organize thousands of food-property measurements and identify a power-law relationship between dielectric constant and loss factor. The system processed 4,451 records across 145 food materials in a
AI.info Team · Written with AI ·
Penn State announced on October 16 that researchers used an AI system to organize thousands of food measurements and identify a relationship between two properties tied to microwave heating. The speed came with a condition: the team compared its AI models against expert-curated data, and the models did not perform equally on every measure. In a paper available online since September 12, the researchers describe AgentXtract, a system that turned scattered scientific results into a searchable database and exposed a power-law link between dielectric constant and loss factor.
Three agents sort papers, measurements and records
Food-property measurements sit across decades of papers, often embedded in tables, graphs, equations and experimental descriptions. AgentXtract assigns different tasks to three large language model agents: screening and interpreting papers, extracting data, and standardizing and assembling records. The workflow focuses on dielectric properties, which describe how food interacts with electromagnetic energy, including the energy used in microwave heating.
The team tested 10 models from OpenAI, Google and Anthropic against a reference dataset verified by human experts. The paper reports that every model achieved precision above 95% and value accuracy above 99%, while recall, API costs and processing times varied. Recall—the share of relevant measurements the system successfully finds—ranged from 79% to 99% in the reported benchmark, showing that accurate extracted values do not guarantee that a system captures every available record.
“All models achieved precision above 95% and value accuracy above 99%, but recall, API costs, and processing times varied.”
Xu Zhou, Yi Zhang, Shyam S. Sablani, Shaojin Wang and Juming Tang, authors of “Agentic AI for food property data: From literature to database to new discovery”
4,451 records produced a pattern across foods
After the comparison, the researchers selected a model to build the full database. AgentXtract assembled 4,451 records for 145 food materials in about 30 minutes, at an API cost of $9, according to the paper. Penn State corresponding author Xu Zhou said gathering and organizing a comparable set manually would have taken an experienced graduate student or postdoctoral researcher six months or more.
Across more than 100 foods, the collected measurements showed a previously unreported mathematical relationship between dielectric constant and loss factor. The paper describes the connection as a power-law relationship. The two properties help characterize how a food responds to electromagnetic energy, a response relevant to microwave processing; the finding emerged when measurements from many sources could be analyzed together rather than one paper at a time.
“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 in Penn State’s announcement. The researchers describe a relationship found in their assembled data, not proof that the same equation applies to every food or every measurement condition.
Human review remains part of the method
The project does not frame the agents as independent scientific judges. Food engineers help identify appropriate sources and assess the reliability of experimental data; the model benchmark itself uses expert-verified records as its reference. That division matters because values reported in papers depend on experimental conditions, and a system can miss data even when the values it does extract are highly accurate.
The database also powers a prototype question-answering system using retrieval-augmented generation, or RAG. A user can ask for dielectric properties of a food such as a potato at a particular frequency; the assistant retrieves relevant database entries, supplies references and explains their connection to microwave heating. The paper presents that system as a way to ground answers in source-linked measurements, rather than relying only on information absorbed during a model’s training.
A narrow case study with a wider data problem
The researchers chose dielectric properties as a demanding test because those values may appear in different formats and must be connected to the food and its experimental conditions. They argue that the same general approach could help organize other scientific data, but the reported results concern this specific food-property dataset. The study appeared online in the Journal of Food Engineering on September 12, ahead of its planned March 2027 issue.
For this case study, AgentXtract’s completed database contains 4,451 records covering 145 food materials, assembled at a reported API cost of $9.