Artificial intelligence is beginning to influence a part of scientific research long considered dependent on human intuition: designing the experiments themselves.

For centuries, progress in physics has relied on researchers asking the right questions and building experimental setups capable of producing clear answers. An international research team now suggests that part of this creative process can be automated. In several areas of physics, AI systems have already proposed experimental configurations that could deliver more precise results than those designed by humans.

Published in Nature, the research outlines how computational methods can explore vast numbers of possible laboratory setups and identify promising combinations that scientists might otherwise overlook.

Finding the best experiment from existing components

A physics laboratory may contain lasers, lenses, mirrors, detectors and electronic components that can be assembled in an almost unlimited number of ways. The challenge is selecting the configuration most likely to reveal a particular physical phenomenon.

Traditionally, this process depends on specialist knowledge, experience and intuition. However, even expert teams can struggle to navigate such a large number of possibilities.

Mario Krenn, now a professor of machine learning in science at the University of Tübingen, encountered this problem while working on a quantum physics experiment as a student in Vienna. His research group could not identify a setup capable of producing the quantum effects they wanted to observe.

Krenn translated the available laboratory components into mathematical descriptions and created an algorithm that could search for viable combinations. After running overnight, the system proposed an experimental setup that met the required conditions, despite the fact that none of the researchers had previously considered it.

An optimization engine, not a chatbot

The technology used to design these experiments is different from generative AI systems such as large language models.

Chatbots learn statistical patterns from large collections of data and generate responses based on those patterns. AI-assisted experiment design is closer to a complex optimization problem. The system searches through an enormous space of possible configurations, simulates their behavior and evaluates how well each one satisfies the researchers’ objectives.

This capability could be particularly valuable in emerging areas where human intuition is still developing.

Philipp Haslinger, head of the Center for Electron Microscopy at TU Wien, points to quantum electron microscopy as one example. Researchers are only beginning to explore concepts such as entanglement in this field. AI could identify microscope designs that humans might never propose, potentially producing better images or enabling entirely new types of measurement.

Similar methods have already been applied to the improvement of fusion reactors, the development of particle detectors and the optimization of gravitational-wave observatories.

When an experiment works but humans cannot explain why

Some AI-generated designs reveal an idea that researchers can quickly understand. Others are more difficult to interpret.

Scientists may be able to calculate and confirm that a proposed configuration performs better without having an intuitive explanation for its success. This creates a familiar challenge in advanced AI: a system can discover an effective solution while leaving humans to reconstruct the reasoning behind it.

Modern computing makes this possible by simulating a wide range of physical conditions in a manageable amount of time. The longer-term ambition is to develop something close to a universal physics simulator. Such a system could use fundamental physical equations to predict what would happen in a proposed experiment and continuously optimize its design.

Humans still need to define the right objective

AI can search the solution space, but researchers must determine what the system is expected to achieve.

Scientists need to define the desired result, the relevant physical constraints and practical limits such as cost, energy consumption, available materials or equipment safety. If the objective is incomplete or poorly formulated, even a highly capable optimization system may generate an impractical or scientifically irrelevant solution.

The role of the researcher is therefore unlikely to disappear. Instead, it moves to a different level. Scientists will spend less time manually evaluating countless configurations and more time framing research questions, defining meaningful objectives and interpreting unexpected results.

This shift resembles previous stages of scientific automation. Researchers no longer perform every calculation by hand, yet mathematical software has not eliminated the need for mathematicians or physicists. It has allowed them to work on more complex problems.

AI-designed experiments could have a similar effect. The technology may expand the range of ideas that science can test, uncover unconventional solutions and accelerate discovery in fields where the number of possible experiments is too large for humans to explore alone.

The next major breakthrough in physics may still begin with a human question. The experimental setup that answers it, however, could be designed by a machine.

Source