A group of Spanish physicists has proposed a machine-learning pipeline that designs the placement and internal optics of gravitational wave detector networks at the same time, rather than fixing a detector geometry first and fitting it to a site afterward. The approach, described in a preprint posted on arXiv on July 28, uses reinforcement learning to choose where interferometer arms go on real terrain and differentiable programming to tune the mirrors inside them.
The method, called IfoScout, is aimed at the design process now underway for next-generation ground-based observatories such as the Einstein Telescope and Cosmic Explorer, and for lunar and space-based proposals. The authors, from the University of Oviedo, argue that today’s process largely separates the two problems: a handful of detector configurations are chosen on scientific grounds, and site integration studies come later. IfoScout combines both into a single optimization loop.
The pipeline works in two stages. First, a reinforcement learning agent, using the Proximal Policy Optimization algorithm, places the arms of the interferometers on a map, rewarded for long arms and penalized for crossing roads, rivers, and excluded perimeters such as settlements and quarries, and for requiring tunnels or bridges. The agent controls the vertex and one end of each L-shaped detector, choosing among 17 possible actions per observatory. A second stage, called AutoGrav, translates the standard LIGO noise calculation software into the PyTorch machine-learning framework so the mirror curvatures and cavity stability can be optimized by gradient descent against the binary neutron star detection range.
To demonstrate the method, the team chose two fictional sites in the Spanish countryside based on real topographic surveys: one near Páramo Leonés, large enough in principle for arms of about 25 kilometers, and a flatter, smaller area near Alcázar de San Juan. The sites are explicitly not under consideration for any real gravitational wave project; the authors chose them as plausible, low-seismicity, sparsely populated terrain.
After training the model for two million steps and testing it from twelve random starting points, the authors present three network configurations that balance length against engineering cost in different ways. The most conservative avoids tunnels entirely for the shorter site and keeps most of the other arms on the surface. The most ambitious reaches a combined arm length of 63 kilometers across the two sites, with the larger site needing 26.33 kilometers of tunnel and 2.48 kilometers of supported structure, at the cost of more underground work. The intermediate configuration lands between the two. Across the three designs, the estimated binary neutron star detection range for the larger site rises from 696 to 1,160 megaparsecs, and for the smaller site from 449 to 564 megaparsecs.
The authors compare the sensitivity curves of their designs with the current Advanced LIGO detectors and the planned A+ upgrade, using the Hanford site’s Newtonian noise model and the original Advanced LIGO design as the starting point. They find their site-adapted designs outperform Advanced LIGO in the 10 to 800 hertz band and A+ in the 10 to 300 hertz band, while stressing that the comparison assumes no future technological improvements and that the results are demonstrations rather than proposals.
The preprint is explicit about its limits. The work is a proof of concept, not a production-ready design tool: the simulations neglect the curvature of the Earth, treat partially underground arms as surface detectors, and omit social factors such as local support, housing, and environmental impact. The authors also note that a sequential optimization of placement and optics can miss coupled effects, since Newtonian and anthropogenic noise depend on exact position, and that merging the two stages into a single differentiable pipeline would require substantial further work. They argue the main value of the method is speed and coverage: the reinforcement learning stage completed its two million steps in under an hour, and each optics optimization took about a minute, allowing many reward structures to be tested quickly.
The code is publicly available on GitHub, and the authors say the pipeline could be used by detector designers as a first step in evaluating candidate sites, including for existing proposals. The broader argument is that systematic, machine-assisted exploration of detector placement could give design teams more confidence in their choices, and open the door to treating the detector and its data-analysis pipeline as one system to optimize together.
Sources
1. arXiv, “Co-design of ground-based gravitational wave detector networks”: https://arxiv.org/abs/2607.27254
2. GitHub, “Daniel-Lanchares/IfoScout”: https://github.com/Daniel-Lanchares/IfoScout