Data on the physical properties of food products, which determine their quality, safety and taste, are scattered across thousands of scientific publications spanning decades. Food engineers can find it difficult to quickly locate and compile the needed figures. Researchers from Pennsylvania State University (USA) proposed solving this problem with artificial intelligence (AI) operating under expert supervision. The work was published in the Journal of Food Engineering (the issue is scheduled for release in March 2027).
The scientists created the AgentXtract system. In it, several AI agents - programs that independently carry out the steps assigned to them - analyze scientific articles together. Importantly, in food engineering valuable experimental data are often hidden not in the text but in tables, charts and formulas. Agents with different "skills" extract numbers from different formats, verify them, bring them to a uniform form and enter them into a database.
To test the system, the researchers tested 10 large language models (these are the algorithms underlying modern chatbots) from OpenAI, Google and Anthropic. They compared accuracy, speed and cost of operation, using data manually verified by experts as the benchmark. According to the head of the work, Xu Zhou, previously such collection took an experienced graduate student or postdoc six months or more. Now more than 4,000 experimental data records were collected and organized in just 30 minutes.
An unexpected result was a new view of the already collected data. The researchers discovered a previously undescribed mathematical relationship between two dielectric characteristics - they show how a product interacts with electromagnetic energy, for example microwaves. These are dielectric permittivity and the dielectric loss tangent (quantities that describe how a material responds to an electric field). This relationship appeared for more than 100 different products. The authors emphasize: they observed a statistical pattern in the collected data, rather than proving a causal mechanism.
In addition, the scientists connected the database to an AI-based question-and-answer system using the RAG (retrieval-augmented generation) scheme. In such a scheme, the model first searches the database for the needed information and then formulates an answer. Testing showed that, relying on human-verified data, the system answers scientific questions more accurately and provides references. For example, to a question about the dielectric properties of potatoes at a certain frequency, it can quickly find records and explain how these properties affect heating in a microwave oven.
Scientists from the University of Washington and Washington State University also took part in the work. The authors note: the proposed approach may be useful not only in food science but also in other fields where it is important to rely on verified experimental data.