A library of 27,000 books, and every one of them is about concrete. Each contains a mix recipe, temperature, and strength. But the information is scattered across paragraphs, tables, and footnotes. For a human, gathering it all together would take years. The program did it in an hour.
The scientists took a large language model, one that can read and understand text, and fine-tuned it to search for concrete data. The model extracted the mix composition, preparation method, and final properties from the papers. For example, how much cement was added, how long it was cured, and what strength was achieved. The collected data went into a table with more than a hundred columns.
Such a database is not just an archive. It can be used to train other programs to predict the properties of concrete for new recipes. The more examples, the more accurate the prediction. And the database is open: any researcher can download it and use it for their own purposes.
Now the library is no longer a maze. It has become a card catalog, where each card is a ready-made recipe. And this works not only for concrete: the same approach can be applied to alloys, polymers, or drugs. Data that used to gather dust in papers is now put to work.
The database contains nearly 9,000 records, and each one has more than 100 characteristics, from cement content to compressive strength.