The load-bearing sentence in the abstract is the framing one. The review treats experimental design as a search for optima over a vast space of hardware configurations subject to practical constraints [3]. Accept that framing and the search algorithm becomes the least interesting component, because an optimizer can query only two things: a simulator that scores a candidate layout, and an objective function that decides what a good score means. Two of the four questions the review organises itself around are exactly those two objects [4], which is to say half the field's open problems concern the map rather than the walking [13].
That matters because the failure mode is asymmetric. A search rewarded by a simulator will find the places where the simulator is wrong and settle there, and a layout that exploits a modelling artefact scores beautifully and cannot be built. The review is candid about this in its own vocabulary, listing trade-offs between computational tractability, experimental feasibility, interpretability and solution reliability [5]. Feasibility appearing on that list is an admission that the winner of the search and the thing you commission are separate questions.
On buildability the supplied material is thinner than the ambition. Melvin, the 2016 photonic framework, is credited with several proposed configurations subsequently realized in laboratories [7], but the annotation gives no count of how many were proposed, so there is no denominator and no build fraction [17]. Tachikoma, from the same year, started with genetic algorithms and later incorporated neural network surrogate models [8], which is the standard speed fix and also a second approximation stacked on the first. PyTheus, in 2023, produced 100 diverse quantum experiments by searching an overcomplete, physics-inspired continuous space [9], seven years on from Melvin [14]. Its output is counted in designs.
The scope has already left the optical bench. The reference list reaches to gradient-based optimization of stellarator coil shapes by an adjoint method [10] and to quasi-isodynamic stellarator configurations put forward as fusion reactor candidates [11]. Coil geometry is not a component you re-tender after the optimizer changes its mind, and that is where the capital question bites: the cheap part of an AI-designed experiment is the search, and the expensive part is owning a simulator trustworthy enough to procure against.
The forward-looking claim is that simulators spanning several physics domains, paired with large suites of experimental objectives, could turn up unorthodox concepts that human intuition struggles to reach [6]. Notice what that puts in the critical path. Someone has to write the suite of objectives, and an objective chosen because it is cheap to evaluate becomes, in effect, the specification of the hardware. The thing this doesn't tell you is whether the objectives used so far are the ones the physicists would defend after seeing what got optimized.
My view, with its condition attached: where a simulator has not been validated against measurement at the tolerance the design leans on, the marginal dollar belongs to the simulator and the objective library rather than to more search compute. Where the physics model is already trusted at that tolerance, the reverse holds and searching harder is the cheap win. The review's statement that discovered configurations often match or exceed human-designed set-ups [2] is a claim scored inside that frame, and reading it as a claim about apparatus already standing in a lab overshoots what the material supports.
