AI proposes physics experiments that beat

Research means asking questions of the universe. For centuries, clever minds have advanced science by devising ingenious experiments designed so their results reveal something about the laws of nature as clearly and unambiguously as possible.

An international research team has now asked: Can this process be automated? Can artificial intelligence develop new ideas for experiments? The answer is a clear yes. In various areas of physics, AI can propose experiments that enable more precise results than experiments designed by humans.

In the journal Nature, the team presented the current state of this new approach to research.

The best experiment from existing components

This is a typical situation in experimental physics: You have a laboratory full of equipment-perhaps lasers, lenses and mirrors, perhaps different detectors and electronic components. All of these can be combined in an almost incomprehensible number of ways. From this vast range of experimental possibilities, you have to select one that can provide new insights into the universe.

Normally, this requires intuition and a great deal of experience. But sometimes even that is not enough, as Mario Krenn discovered. Today, he is a professor of machine learning in science at the University of Tübingen. As a student in Vienna, he was working on the setup for a quantum experiment. But neither he nor the other members of his research group could find a suitable experimental configuration capable of demonstrating the desired quantum effects.

So Krenn decided to ask the computer. He described the individual components available to him mathematically, then had an algorithm search for combinations of these components that would result in a meaningful experiment.

"Programming it only took a few hours. Then I went home and left the computer running," Krenn says. "When I came into the office the next day, the program had produced a file containing a proposed solution. Of course, that was extremely exciting."

A search problem, not a chatbot

This approach has little in common with the kind of AI familiar from large language models. Chatbots are trained on enormous amounts of data and then generate solutions that are statistically likely. When searching for new physics experiments, the task is entirely different.

"It is an enormous optimization problem," Krenn says. "There is an overwhelmingly large space of possible experiments that can be built from the available components."

"The results are impressive," says Philipp Haslinger, head of the Center for Electron Microscopy at TU Wien. "In electron microscopy in particular, we are only now beginning to work systematically with entanglement and new quantum-mechanical microscopy concepts."

This approach has already been used to improve fusion reactors, develop new ideas for particle detectors and generate proposals for making gravitational-wave detector systems even more sensitive.

"Sometimes you look at these computer-generated experimental proposals and quickly understand the idea behind them-why the new concept works better than previous approaches," Krenn says. "But sometimes it is also very difficult to understand."

This is possible because modern computers can simulate a wide range of physical situations within a manageable amount of time.

"The goal is to develop something like a universal physics simulator," Krenn says. "Today, a few important fundamental equations of physics can already take you a very long way."

The art of defining the goal

Defining that objective, however, remains the task of humans.

"That is precisely the challenge: defining as accurately as possible what you actually want, and which constraints have to be satisfied-for example, a maximum cost or a maximum amount of energy the device can absorb without exploding."

But isn't it also a little unfortunate if, in the future, we leave the great eureka moments of science to computers?

"No, absolutely not," Krenn says. "Human work is simply shifting to a higher level."

Publication details

Jonathan Klimesch et al, Designing physics experiments with artificial intelligence, Nature (2026). DOI: 10.1038/s41586-026-10898-6

**Journal information:** [Nature](https://phys.org/journals/nature/)

Key concepts

laboratory experiments Numerical techniques

Provided by Vienna University of Technology