What if AI could not only help analyze scientific data, but also come up with entirely new experiments? An international research team has demonstrated that this is already possible: algorithms can systematically search through enormous numbers of possible experimental configurations and identify setups that can deliver more precise results or enable new measurement capabilities.
Mario Krenn, Professor of Machine Learning in Science at the University of Tübingen, a member of the Cluster of Excellence: New Perspectives for Science and of the Tübingen AI Center, has been working on this approach for several years. In an early quantum experiment, an algorithm searched through combinations of available components and found an experimental setup that the researchers themselves had been unable to devise. “It’s a massive optimization problem,” says Krenn. “There is an unimaginably vast space of possible experiments that can be set up with the available components. And the computer must systematically search this space to find the best possible solution.”
Today, AI-driven experimental design is being explored in areas ranging from electron microscopy and fusion research to particle detectors and gravitational-wave detection. For Krenn, this does not mean handing scientific discovery over to machines. Instead, AI could give researchers new tools to tackle problems beyond the reach of human intuition alone.
The paper has just been published by Mario Krenn's group together with their partners from the TU Wien and other co-authors:
Klimesch, J., Arlt, S., Ruiz-Gonzalez, C. et al. Designing physics experiments with artificial intelligence. Nature 657, 47–58 (2026). https://doi.org/10.1038/s41586-026-10898-6
Read the full story on the website of TU Wien.