Artificial intelligences are no longer content to assist physicists in their experiments: they are proposing innovative setups themselves.
A review published on September 2 in Nature takes stock of this development. The authors examined work already carried out in several branches of physics. Their finding is precise: some AI tools now explore the very shape of an experiment, instead of merely modifying a few settings decided in advance by humans.
To understand the difference, consider a setup made up of sensors, mirrors, or other elements. A conventional method may search for the best values among several parameters. A more ambitious approach can also decide which elements to use, how to connect them, and where to place them. The number of possibilities then becomes immense.
This is where artificial intelligence comes in. Researchers first define what the future device must accomplish and the constraints it must meet. A program then explores numerous configurations, generally using simulations. It can thus arrive at an arrangement that human intuition would not have found.
This approach has already been applied to quantum optics experiments, particle detectors, neutrino antennas, and gravitational-wave detectors. In several studies reviewed by the authors, the resulting configurations match or exceed the performance of models designed by specialists. Some proposals were subsequently built in the laboratory.
Humans therefore do not disappear from the process. They choose the scientific question, set the material limits, and decide how to measure the quality of a solution. An excellent configuration on paper is of little interest if it is impossible to manufacture or too sensitive to real-world imperfections.
Another difficulty is that a machine may find a high-performing setup without providing an explanation of how it works. Physicists must then understand why this arrangement works. This step can itself be instructive, because an unusual solution may reveal a different way of exploiting already known physical laws.