Ahead of their time: Restricted to pre-2018 research, AI flags hydrogen catalyst reported in later experiments

Researchers have used artificial intelligence to identify promising catalysts using only the scientific knowledge that was available years before their actual discovery. Their findings could accelerate the search for more efficient hydrogen production and other valuable chemical processes.

Predicting the future is an imprecise science. The output can only be theoretical because the future hasn't yet happened. But what if you could test a prediction against a future that has already occurred? Researchers from the ARC Centre of Excellence for Carbon Science and Innovation have found an ingenious way to do just that.

In effect, to move forward, they turned back the clock: They restricted the scientific literature used by their system to information published before a particular date, then asked it to identify promising electrocatalysts.

The results suggest that some of tomorrow's scientific breakthroughs may already be hiding in the enormous body of research that has already been published.

Millions of possibilities

The research, published in ACS Catalysis, explores how large language models (LLMs) can help scientists identify promising electrocatalysts—materials that accelerate chemical reactions involving electricity.

The research team, comprising center members Dr. Yiran Jiao, professor Yan Jiao and professor Shi-Zhang Qiao, together with collaborators from computer science, focused on carbon-supported high-entropy alloys (HEAs). These materials typically consist of five or more different metals in relatively similar proportions, combined with carbon-based supports.

The interaction between multiple metals and carbon-based materials can give rise to unique properties, making them promising candidates for applications such as hydrogen production.

But finding the right combination is an enormous challenge. According to Yiran Jiao, once scientists account for different elements, their proportions and the structures they form, the number of possible HEA combinations can run into the millions. This makes testing every possibility impractical.

To combat this, researchers have traditionally relied on theoretical calculations, existing scientific knowledge and trial and error through experimentation to identify promising candidates.

Each approach comes with limitations. Computer simulations can test large numbers of materials, but their simulated conditions do not always reflect what happens in a laboratory. Experimental testing is more reliable, but synthesizing and evaluating new materials takes considerable time and effort.

Researchers also tend to favor combinations of elements that are already known to work well together, which can mean less obvious combinations with potentially valuable properties are overlooked.

Finding tomorrow in yesterday

The research team set out to find another way. Their approach used large language models to gather information from scientific publications, then analyzed that information to identify patterns and relationships that could point toward promising new materials.

"There's a large amount of information in research publications," says Yiran Jiao. "You can't go through them one by one to extract all the knowledge. That's what large language models can help us do."

But Yiran Jiao estimates that reviewing and extracting information from just 100 scientific papers can take a researcher several hours. In contrast, the team's new research approach can extract information from thousands of publications within a similar time frame.

"We were trying to reduce the possible candidates from millions of combinations to perhaps 10 or 20," he explains. "That could greatly reduce the time and effort needed to investigate them."

But extracting information was only the first step. The researchers also developed a new calculation method to analyze the patterns and relationships in that information. This allowed them to predict which combinations of metals might produce catalysts with greater stability and activity, including combinations that had never been tested experimentally.

To test the approach, the researchers restricted the system to scientific knowledge available up to 2017, then compared its predictions with research published subsequently. In one example, the system identified a promising catalyst combining chromium, cobalt, iron, manganese and nickel. The alloy, known as CrCoFeMnNi, was subsequently reported experimentally as one of the pioneering examples of high-entropy alloys applied to hydrogen electrocatalysis.

"We still need humans to examine AI's output because it isn't always reliable," Yan Jiao says. Large language models can generate inaccurate information or suggest things that do not exist. To address this, the researchers manually assessed 100 examples of the AI's output, identifying correct and incorrect results. They then used these examples to train and refine the system, improving its accuracy.

With this relatively small amount of human input, the refined system was able to extract scientifically meaningful information from more than 5,000 research papers, which demonstrates how human expertise can improve AI's ability to analyze scientific knowledge on a much larger scale.

Why this matters for hydrogen production

Producing hydrogen by splitting water requires electricity, and electrocatalysts help the chemical reactions proceed efficiently. Finding catalysts that are more active and stable could improve the performance of hydrogen production technologies, which could potentially reduce costs and make the process more commercially attractive.

By narrowing millions of possible material combinations to a small number of promising candidates, the researchers' approach could significantly reduce the time and resources needed to investigate new materials. However, the study involved computational predictions rather than new experimental testing.

The team has therefore not demonstrated improvements in industrial hydrogen production or quantified potential cost savings. The approach is intended to guide scientific investigation, not replace the experimental work needed to establish whether a material performs as predicted.

Beyond hydrogen

Hydrogen may just be the beginning—the underlying approach could potentially be applied to a much wider range of materials and electrochemical reactions.

One possibility is carbon dioxide reduction, in which CO₂ is converted into useful chemical products. More effective catalysts could make these processes more efficient and commercially attractive, potentially creating new opportunities to turn captured CO₂ into valuable chemical products. Yiran Jiao says the team is also interested in an even more ambitious possibility: whether artificial intelligence could help scientists identify entirely new chemical reactions, rather than simply finding better materials for reactions that are already well understood.

A future of discovery

Perhaps the most intriguing outcome of the research is what remains unknown. Having demonstrated that their approach could identify promising materials subsequently reported by other scientists, the team has also recommended several previously unexplored electrocatalyst candidates for research. The system predicts that these materials could offer improved stability and activity for hydrogen production, potentially outperforming existing candidates.

But those predictions have yet to be confirmed experimentally. That is now an opportunity for experimental researchers: to take the candidates identified by AI, synthesize them and discover whether they perform as expected.

The results the team has produced so far suggest that valuable discoveries may already be waiting within the scientific knowledge we possess. The next test is whether that same knowledge can help scientists find something genuinely new.