By 2017, an artificial intelligence system had already picked out a catalyst that researchers would later report in experiments: an alloy made from chromium, cobalt, iron, manganese and nickel. The prediction came from a deliberately time-limited search through scientific papers, offering a way to test whether AI can find useful clues in research published before a discovery is known.

The work, published in ACS Catalysis, was carried out by researchers from the ARC Centre of Excellence for Carbon Science and Innovation and computer science collaborators. The team focused on electrocatalysts, materials that speed up chemical reactions involving electricity. Better catalysts could help make hydrogen production more efficient, among other uses.

The researchers studied carbon-supported high-entropy alloys. These alloys typically contain five or more metals in roughly similar amounts, combined with a carbon-based support. Their mix of elements can create useful properties, but the number of possible combinations quickly grows into the millions when scientists vary the metals, their proportions and the structures they form. Testing every option in a lab is not practical.

To narrow the search, the team used large language models, or LLMs, to pull information from scientific publications and identify patterns between materials and their properties. Team member Yiran Jiao said that extracting knowledge from just 100 papers can take a researcher several hours. The team’s approach can process information from thousands of publications in a similar time frame, according to Jiao.

The system then used a calculation method to assess those patterns and predict metal combinations that might offer greater stability and activity. The goal, Jiao explained, was to reduce millions of possible candidates to perhaps 10 or 20 worth closer investigation. That would not replace laboratory experiments, but it could help researchers decide where to spend their time and resources.

The team tested its approach by limiting the literature available to the system to research published up to 2017. It then compared the system’s predictions with studies that appeared later. Among the combinations it identified was CrCoFeMnNi, an alloy of chromium, cobalt, iron, manganese and nickel. Later research reported the alloy experimentally as one of the pioneering examples of a high-entropy alloy used for hydrogen electrocatalysis.

The result is a useful test of a familiar promise: AI might help scientists search a vast body of knowledge for patterns that people could miss or take too long to find. It does not show that every prediction will work in practice, and experiments remain essential. But by showing that information available before 2018 could point toward a material reported later, the study suggests that scientific papers may contain leads to discoveries still waiting to be recognized. The next step is to see how well this approach can guide the search for other catalysts and chemical materials.