In the early days of applying artificial intelligence (AI) for science, catalysis researchers used AI to simulate reactions, analyze microscopy images, and scour the literature for new results. Across these independent efforts, there was no consensus view in the field.
John Kitchin, Núria López, Neil Schweitzer, and Hongliang Xin wanted to share learning within their research community and offer a vision for the future. “If we want to translate these isolated proofs-of-concept into societal impact, then we need to align investments and standards from academia, national labs, industry, and funding agencies,” says Xin, professor of chemical engineering at Virginia Tech.
In Nature Catalysis, Xin, Kitchin, López, Schweitzer, and their co-authors propose a roadmap for integrating AI into catalysis in order to accelerate the design, discovery and optimization of new catalytic materials and processes.
“AI is the next logical step in the evolution of science,” says Kitchin, professor of chemical engineering at Carnegie Mellon. Take the last several decades. Computers enabled us to do things that we couldn’t do with paper and pencil or even with calculators. Machine learning was a step up in building more flexible and data-driven models to make predictions and identify patterns that we weren’t able to see before.
With AI, we now have large language models with which we can use natural language, instead of only numbers and equations. “This continuous evolution brings us to the newest capability, an iterative loop in which AI is able to do things with machine learning, see what happens, and then do new things,” says Kitchin.
The path outlined by Kitchin, López, Schweitzer, Xin, and their co-authors points toward AI-empowered human-machine collaboration in the field of catalysis. “AI can scale, automate, and optimize experiments, in silico or physical, while humans provide constraints, judgement, and scientific direction,” says Xin.
Researchers see the potential for AI to help meet the urgent need for new catalytic materials in the chemical, energy, and environmental industries. Energy demands and waste streams are surging globally. New catalysts are a key to sustainable manufacturing, renewable energy solutions, and climate resilience.
“Applying AI to create more sustainable industrial processes requires complex chemistry, physics, and engineering. Heterogeneous catalysis encompasses multiscale and multiphysics phenomena, and so AI needs datasets that capture that complexity across scales,” says López, group leader at the Institute of Chemical Research of Catalonia.
López and co-authors call for standardizing catalysis data with shared frameworks and a common vocabulary for humans and machines. This requires combining experimental and computational data. Digital infrastructure and cloud platforms offer the means to automate data capture and to integrate data storage, access, and AI workflows.
The field of heterogeneous catalysis currently lacks enough data to meet the demands of AI. “It’s been more than 100 years since there was a massive centralized effort to generate data on catalysts,” says López. That resulted in the Haber-Bosch process, used for the industrial production of ammonia. “Since then, industrial developments have been scattered in academia or protected by companies. We need a robust data ecosystem to steer the integration of AI in catalytic process development.”
López, Kitchin, Schweitzer, Xin, and their co-authors also advocate the simultaneous development of multimodal foundation models. These models combine information from different data types: structures, spectra, images, and text. General-purpose models trained on domain knowledge can be fine-tuned for specific tasks.
Multimodal foundation models are needed in the agentic labs that the co-authors envision for the future. In these labs, AI doesn’t replace scientists; it empowers them.
When routine work, such as writing code, running instruments, and analyzing data, is offloaded to AI, scientists will be free to focus on critical thinking. In an agentic future lab, an AI agent will propose what to do, direct robotic systems to execute and analyze experiments, and iterate based on the results. A human scientist can integrate themself at any point in the loop, to approve or change the proposed actions.
This vision is within reach. “The field is advancing so quickly. What ‘agentic catalysis’ meant in 2024 and what it means today are quite different,” says Kitchin. “In 2024, we could use ChatGPT or Copilot to generate answers. Those tools can now interact with laboratory instruments and with your computer in a continuous loop. The ability to couple a large language model with an actual physical device creates an auditable, verifiable system.”
Kitchin and co-authors propose an open and remotely-accessible network of agentic future labs. To prepare, the catalysis field will need new approaches to training scientists, building infrastructure, and coordinating across institutions, disciplines, and industrial sectors.
“Imagine AI scientists that can perceive inputs, reason, plan, and take actions,” says Xin. “Working alongside them, it will be possible to make scientific discoveries beyond our human capabilities.”
Reference publication
Roadmap for transforming heterogeneous catalysis with artificial intelligence
Xin, H.; Kitchin, J. R.; López, N.; Schweitzer, N. M.; Artrith, N.; Che, F.; Grabow, L. C.; Gunasooriya, G. T. K. K.; Kulik, H. J.; Laino, T.; Li, H.; Linic, S.; Medford, A. J.; Meyer, R. J.; Peng, J.; Phillips, C.; Qian, J.; Qi, L.; Shaw, W. J.; Ulissi, Z. W.; Wang, S.; Wang, X.
Nat Catal 2026
DOI: 10.1038/s41929-026-01479-x
La entrada AI in the catalysis labs of the future se publicó primero en ICIQ.