A team at the University of Rochester has cracked a major barrier in AI-powered materials discovery: instead of producing complex numerical data that only experts can decipher, their new method generates step-by-step instructions that any researcher can understand and execute. The breakthrough uses large language models (LLMs) like ChatGPT to guide the discovery of new catalysts for converting carbon dioxide into fuel, potentially accelerating a process that currently takes a decade or longer .

Why Can't Researchers Just Use Traditional AI for Materials Discovery?

For years, the standard approach to AI-powered materials discovery relied on a technique called Bayesian optimization, which efficiently explores vast experimental design spaces. The problem: it produces dense numerical outputs about a material's molecular structure that require deep expertise to interpret and act upon. This technical barrier has kept many researchers from using AI tools, even when they could accelerate their work significantly .

The University of Rochester team, led by Marc Porosoff, an associate professor in the Department of Chemical and Sustainability Engineering, recognized this gap and reimagined the entire workflow. Instead of asking AI to optimize numerical parameters, they asked it to do something more intuitive: describe how to make a material, just like a recipe .

"We're able to leverage the pre-trained knowledge of large language models and well-established statistical methods for materials discovery to help us as researchers navigate large experimental design spaces more efficiently," said Marc Porosoff.

Marc Porosoff, Associate Professor, Department of Chemical and Sustainability Engineering, University of Rochester

Porosoff uses a helpful analogy to explain the difference. Someone could describe a cup of coffee by its taste, color, and aroma, or by specifying the bean type, grind size, brewing apparatus, and water temperature. Both describe the same coffee, but the second approach gives you a reproducible recipe that others can follow .

How Does the New LLM-Based Method Actually Work?

• Natural Language Input: Researchers describe the material they want to create using plain English, specifying desired properties or applications like "a catalyst for converting CO2 to methanol."
• AI-Generated Procedures: The LLM produces a set of concrete experimental steps that researchers can understand and execute in their labs without needing to interpret complex mathematical models.
• Iterative Feedback Loop: As researchers run experiments and observe results, they feed the outcomes back into the AI model, which refines its suggestions and continues iterating until the team reaches their goal.

This approach dramatically reduces the expertise barrier. Shane Michtavy, a chemical engineering PhD student at the University of Rochester who helped develop the method, explained that using pre-trained LLMs allows researchers to explore with less data than traditional models require, since the language models come pre-loaded with built-in knowledge about the physical world and catalysis .

"Our method reduces the technical barrier associated with using Bayesian optimization, which is a well-established method for efficiently exploring large and complicated parameter spaces. Using pre-trained LLMs allows users to explore using less data than traditional models, as they are deployed in a frozen state with built-in knowledge of the physical world and catalysis," explained Shane Michtavy.

Shane Michtavy, Chemical Engineering PhD Student, University of Rochester

What Results Did the Researchers Actually Achieve?

The proof of concept is striking. In one experiment, the team aimed to identify catalysts for converting carbon dioxide and hydrogen into carbon monoxide and water using trimetallic catalysts, which are made from three different metals. Theoretically, there were approximately 360,000 possible experiments they could have run to find an ideal candidate .

Using their LLM-based method, they found an ideal catalyst in just ten experiments. That represents a reduction of roughly 99.997% in the experimental search space. The study, published in ACS Central Science, demonstrated the method across several live experiments, proving it works as more than just a theoretical concept .

The success caught the attention of the U.S. Department of Energy. The Advanced Research Projects Agency-Energy (ARPA-E) announced nearly $3 million in funding to scale the approach through its Catalytic Application Testing for Accelerated Learning Chemistries via High-throughput Experimentation and Modeling Efficiently (CATALCHEM-E) program .

What's the Bigger Vision for This Technology?

The immediate goal is ambitious but focused: Porosoff's team will use the ARPA-E funding to develop catalysts that convert carbon dioxide and hydrogen into methanol and ethanol, which are valuable fuels and chemical feedstocks. Ethanol, for instance, is a key additive in gasoline and is used in pharmaceuticals, cosmetics, and numerous other applications .

The longer-term vision is even more transformative. Right now, moving from a new catalyst concept to laboratory testing to deployment in an industrial reactor typically takes a decade or more. Porosoff stated that the CATALCHEM-E program aims to compress that timeline by an order of magnitude, reducing it to roughly a single year. He believes AI with text-based representations will be a major factor in achieving that acceleration .

"Right now, it takes a decade or longer to go from conceptualizing a new catalyst to testing it in a lab to putting it in a real reactor. The CATALCHEM-E program aims to cut that by an order of magnitude to a single year, and we think using AI with text-based representations will be a big factor in shortening the development cycle," said Porosoff.

Marc Porosoff, Associate Professor, Department of Chemical and Sustainability Engineering, University of Rochester

The project, which began in July 2024 and runs through 2029, brings together a multi-institution team including Virginia Polytechnic Institute and State University, Stanford University, Northwestern University, A*STAR Institute of Sustainability for Chemicals, Energy and Environment (ISCE2) in Singapore, and OxEon Energy, a small business based in Salt Lake City .

Why Does This Matter Beyond Chemistry Labs?

The breakthrough illustrates a broader shift in how AI is being applied to the physical world. Rather than replacing human expertise, the new method amplifies it by removing technical barriers and making AI tools accessible to researchers who lack deep machine learning backgrounds. This democratization of AI-powered discovery could accelerate innovation across materials science, chemistry, and related fields .

The work was supported by funding from the National Science Foundation, the National Institutes of Health, and the U.S. Department of Energy, reflecting the significance researchers and policymakers place on accelerating materials discovery for energy applications .