Four Problems Stand Between AI and a Machine-Designed Experiment

The hardest part of a physics experiment is often deciding what it should look like: which optical parts, in which order, how long the arms are, where the detector sits. That judgment is taught by apprenticeship, and until recently software's role in it was to tune a handful of numbers a physicist had already chosen.

A review published in Nature on 2 September argues that the role has grown. Thirteen researchers wrote it, led by Jonathan Klimesch and Mario Krenn of the University of Tübingen. Between them they work in quantum optics, gravitational-wave physics, neutrino astronomy and particle physics, and they describe a field in which software proposes whole experimental layouts rather than settings for an existing one. A review reports no new measurement of its own; it is an argument about where a field has arrived. This one puts its central claim in a single sentence: the designs that come out "often challenge established design conventions while matching or even exceeding the performance of human-designed set-ups," the authors write.

The field is younger than the phrase "AI-designed experiment" suggests. The earliest work the review reaches back to dates from 2016. One is Melvin, a program from Krenn and colleagues that searched for photonic quantum experiments: setups made of lasers, crystals and mirrors. The other is Tachikoma, from P. A. Knott, aimed at designing experiments that measure quantum systems as precisely as possible. The review's own note on the Melvin paper records that several of its proposed configurations were "subsequently realized in laboratories." That detail is the point where the calculations turn into experiments.

Around that decade of work the authors organize the field around four questions. Each is an engineering problem rather than a philosophical one.

The first is the search space: the set of components and arrangements the program is allowed to consider at all. Make it too narrow and the software can only rediscover what physicists already build; make it too wide and no search will ever finish.

The second is the simulator, the software stand-in for a real optical bench. It has to run fast enough to be called over and over, and be faithful enough that a layout that wins on screen still wins on the table.

The third is the objective: turning a scientific goal into a single number the program can push up. "A better gravitational-wave detector" has to become a score, and the score decides what comes back.

The fourth is the search itself, over choices of two different kinds at once. Whether to include a particular component is yes or no; where to put it, and at what angle, is a dial that turns smoothly.

The work gathered under those headings is spread across physics rather than confined to one corner of it. Programs that breed and mutate candidate designs have shaped antennas for catching the highest-energy neutrinos. Another family of search methods tuned a detector component for the planned Electron-Ion Collider. A third search took on the layout of laser interferometers, the L-shaped machines that detect gravitational waves. Krenn ran it with Yehonathan Drori of Tel Aviv University and Rana Adhikari of Caltech, reporting designs that outperform the next generation now being planned. A similar tool, XLuminA, does the same job for high-resolution microscopes. Krenn is also an author of the Melvin paper, the interferometer search and XLuminA, among other work discussed in the review, and one of its two corresponding authors. The competing-interests statement notes that Krenn, Klimesch and Arlt are founding Feyer GmbH, an AI-for-invention company, and have been selected for funding through the SPRIND Next Frontier AI Challenge. The declaration says this activity began after the manuscript was submitted and that neither the funding nor the company’s activities contributed to the review. The review’s case for AI-designed experiments thus comes partly from researchers who are also building a company around machine-driven invention.

However, this remains a forward-looking projection rather than an accomplished milestone. Ultimately, the validity of these concepts will face the same practical test that Melvin's designs met a decade ago: physical implementation on a laboratory bench.