Artificial intelligence is capable of finding promising materials based only on scientific data available many years ago. This was shown by staff at the ARC Centre of Excellence for Carbon Science and Innovation, an Australian center for carbon science and innovation. The results were published in the journal ACS Catalysis.
The scientists set up an unusual experiment: they restricted the system to literature published before a certain date and asked it to propose promising electrocatalysts, materials that accelerate chemical reactions involving electricity. The predictions were then compared with what was published later.
The focus was on high-entropy alloys on a carbon support. These are materials in which five or more different metals are mixed in roughly equal proportions and combined with a carbon base. Such combinations can acquire unusual properties useful, for example, for producing hydrogen.
However, selecting the right combination is extremely difficult. According to one of the authors, Dr. Yiran Jiao, taking into account different elements, their proportions and structures, the number of possible variants reaches millions. Testing each one is practically impossible.
Traditionally, scientists rely on theoretical calculations, known data and trial and error. But each approach has limitations: computer modeling does not always reflect real laboratory conditions, and experiments require a lot of time and resources. In addition, researchers often choose combinations of elements that have already proven themselves well, missing less obvious options.
The new approach uses large language models, artificial intelligence systems trained to analyze texts. They extract information from thousands of scientific publications and identify patterns pointing to promising materials. As Yiran Jiao notes, manually processing even a hundred articles is a task that takes several hours, while the system handles thousands in a comparable time.
The researchers then developed a method for analyzing the extracted data to predict which combinations of metals would yield more stable and active catalysts. To test the approach, the system was limited to knowledge up to 2017. Among other things, it pointed to an alloy of chromium, cobalt, iron, manganese and nickel (CrCoFeMnNi). This material was later indeed described in an experimental paper as one of the first examples of high-entropy alloys for electrocatalysis in hydrogen technologies.
The authors emphasize that humans must still oversee AI results. Large language models sometimes produce inaccurate information or suggest something that does not exist. To improve reliability, the researchers manually checked 100 examples of output, marked correct and incorrect answers, and used them to further train the system. After that, the improved system extracted scientifically significant information from more than 5,000 articles.
The work is important for hydrogen production: electrocatalysts help split water more efficiently. However, the study is based on computer predictions rather than new experiments. The authors did not demonstrate improvements in industrial hydrogen production and did not assess possible savings. The approach is intended to guide scientific search, not replace laboratory verification.
Potentially the method is applicable more broadly, for example to carbon dioxide reduction reactions, where CO2 is converted into useful chemical products. In addition, the scientists suggest that AI could also help search for entirely new chemical reactions, not just improve materials for already known ones. The system also proposed several previously unstudied electrocatalyst candidates that, according to its forecast, may outperform existing ones. But these predictions are still awaiting experimental confirmation.