Researchers Use AI to Cut Materials Discovery Time From Years to Weeks
Scientists at the University of Rochester have developed a method using large language models to accelerate the search for new catalysts, reducing thousands of possible experiments to a handful of actual tests. In one case, researchers narrowed 360,000 possible experiments down to 10.
The team, led by Marc Porosoff, an associate professor in chemical and sustainability engineering, published their findings in ACS Central Science. The method allows researchers to describe materials they want to create using natural language, then receive step-by-step instructions for experiments to produce them.
Traditional AI approaches for materials discovery produce complex numerical data that requires deep expertise to interpret. The new method instead generates procedures that any trained researcher can understand and execute, then feed results back into the model to refine the approach.
How It Works in Practice
Porosoff compares the approach to describing coffee. Someone could describe it by taste, color, and aroma-or by specifying bean type, grind size, apparatus, and water temperature. The second approach provides a recipe others can replicate.
The team applied this principle to catalysts for energy applications. They used language-based descriptions to specify not just what materials should do, but the exact steps needed to create them.
In a live experiment, researchers sought catalysts to convert carbon dioxide and hydrogen into carbon monoxide and water using trimetallic catalysts made from low-cost metals. The AI model identified an effective candidate in 10 experiments instead of the estimated 360,000 possibilities.
Scaling Up With Federal Funding
The U.S. Department of Energy's Advanced Research Projects Agency-Energy (ARPA-E) announced nearly $3 million in funding to expand the method. The money will support a multi-institution team working to create catalysts that produce methanol and ethanol from carbon dioxide and hydrogen.
The project includes researchers from the University of Rochester, Virginia Tech, Stanford University, Northwestern University, A*STAR Institute of Sustainability for Chemicals, Energy and Environment in Singapore, and OxEon Energy, a Salt Lake City-based company.
The work is scheduled to begin in July and run through 2029. Current timelines for developing new catalysts stretch a decade or longer from concept to real-world deployment. ARPA-E's program aims to compress that to a single year.
"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," Porosoff said. "We think using AI with text-based representations will be a big factor in shortening the development cycle."
Why This Matters for Researchers
Large language models arrive with built-in knowledge of the physical world and catalysis principles. This means researchers can explore possibilities using less data than traditional machine learning models require, reducing the technical barrier to entry.
Shane Michtavy, a chemical engineering PhD student who helped develop the method, noted that the approach works well with complex materials like trimetallic catalysts. "Our method reduces the technical barrier associated with using Bayesian optimization," he said.
The funding supports work on converting carbon dioxide to methanol first, then extending to higher alcohols like ethanol. Ethanol serves as a fuel additive and ingredient in pharmaceuticals and cosmetics. The team ultimately aims to deploy the model commercially so industries can create their own catalysts.
Learn more about Generative AI and LLM applications and AI for Science & Research.
Your membership also unlocks:
