AI can design viable physics experiments, Nature study shows—but scientists still set the rules

An international research team has demonstrated that artificial intelligence can propose workable physics experiments by searching combinations of available components under constraints defined by scientists.

The peer-reviewed study, published in Nature on September 2, covers examples in quantum optics, electron microscopy, fusion research, particle detectors and gravitational-wave detection. The system is better understood as a powerful optimiser than as a chatbot independently deciding what science to pursue.

How the method works

Researchers provide a goal, a catalogue of components and physical or engineering constraints. The algorithm searches a very large design space for arrangements that maximise the chosen objective. That can expose unconventional combinations a person might not test manually.

The paper argues that a shared design framework can be used across several branches of physics even though the hardware differs. A detector layout and an optical table are not the same problem, but both can be represented as constrained optimisation.

What the result does not mean

The AI does not replace experimental judgement. Scientists still decide whether the objective is meaningful, whether the simulation captures relevant physics, whether components are realistically available and whether a proposed setup can be operated safely.

A design that succeeds in software can fail because of calibration, noise, manufacturing tolerances or an omitted constraint. Proposed experiments therefore need simulation checks, expert review and physical replication before their results can support a scientific claim.

Sources: peer-reviewed Nature paper; TU Wien research summary.

Illustration accompanying the research release. Image: TU Wien.