Pith. sign in

REVIEW 3 major objections 5 minor 64 references

IfoScout combines reinforcement learning and differentiable optics to co-design detector networks on real terrain, achieving BNS ranges up to 1160 Mpc.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A proof-of-concept ML pipeline co-designs two-site gravitational-wave detector networks, producing three example configurations with 15–44 km arms and BNS ranges up to 1160 Mpc.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A genuinely new two-stage ML pipeline for GW detector siting, but the sensitivity numbers rest on an unvalidated differentiable surrogate. the 3 major comments →

arxiv 2607.27254 v1 pith:5GOKDQRJ submitted 2026-07-28 astro-ph.IM gr-qc

Co-design of ground-based gravitational wave detector networks

Daniel Lanchares , Lysiane Mornas , Luigi Toffolatti , Pietro Vischia This is my paper
classification astro-ph.IM gr-qc
keywords co-designgravitational-wave detectorsreinforcement learningdifferential programmingdetector site selectionsensitivity optimizationbinary-neutron-star rangeinterferometer design
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

IfoScout is a two-stage machine-learning pipeline that co-designs the layout and internal optics of ground-based gravitational-wave detector networks from topographic data. The first stage uses reinforcement learning to place the vertex and end of an L-shaped interferometer on a real elevation map, rewarding long arms and a favorable network orientation while penalizing crossings of roads, rivers, and settlements. The second stage uses gradient-based optimization of a differentiable noise and range model to tune arm length and mirror curvatures, enforcing a stable optical cavity and maximizing the binary-neutron-star (BNS) detection range. On two plausible Spanish sites the pipeline produced three two-detector configurations with arm lengths from 24 to 44 km, tunnel lengths up to 26 km, and BNS ranges up to 1160 Mpc, outperforming the baseline current-generation design in the 10–800 Hz band. The paper's case is that such fast, flexible co-design makes site-specific optimization practical for future detector studies.

Core claim

The central claim is that detector placement and internal design need not be sequential: a single pipeline can simultaneously choose where to put an interferometer and how to configure its optics. IfoScout demonstrates this by training a reinforcement-learning agent (GWymnasium) to pick vertex and end positions for two L-shaped detectors on real digital elevation maps, with a reward that includes arm length, a 45-degree relative-orientation term, and penalties for tunnels, bridges, and excluded areas. The best configurations are then passed to AutoGrav, a differentiable re-implementation of standard noise and range calculations, which optimizes arm length and mirror curvature radii under a c

What carries the argument

The load-bearing machinery is the two-stage loop. GWymnasium is a reinforcement-learning environment that represents each L-shaped interferometer by a vertex point and an end point; the agent can move each point north, south, east, or west with fine or coarse steps, or stay still, and receives a reward that adds arm length, favors a network-wide relative orientation near 45 degrees, and subtracts penalties for roads, rivers, tunnels, supported sections, and proximity to site edges. AutoGrav is a differentiable re-implementation of the standard gravitational-wave noise calculator and BNS-range estimator; it computes gradients of the range with respect to arm length and the two mirror curvatur

Load-bearing premise

The optimization's ranking of configurations is reliable only if the BNS range computed from a simplified, differentiable noise model (using a current-generation detector's newtonian noise and design parameters) accurately reflects real sensitivity at a new site, and if the elevation-based terrain costs capture true civil-engineering difficulty.

What would settle it

Recompute the BNS range for one of the optimized configurations (e.g., C3-S1) with a full, non-differentiable noise model that uses the actual site's newtonian and anthropogenic noise plus the real topography, and compare it with the AutoGrav prediction; a mismatch larger than the optimization tolerance would show the pipeline's sensitivity estimates and length-versus-cost trade-offs are not trustworthy.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Project teams could use the pipeline to systematically search many site-specific configurations, including detectors longer than the 15–20 km surface arms often considered infeasible in Europe.
  • The same method, with 3D geological data replacing 2D elevation maps, could be used to site and design underground detectors.
  • Because the RL training takes under an hour and each optics optimization about a minute, designers can rapidly test many reward functions encoding different priorities, from cost savings to scientific reach.
  • If the computed sensitivities hold up, site-adapted designs can beat current-generation detectors across 10–800 Hz without any new technology, purely through co-optimized geometry and optics.
  • The pipeline can be paired with data-analysis software to move toward full hardware–software co-design of future detectors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same co-design loop could be applied to existing detector sites to evaluate alternative configurations or upgrade paths, not just to greenfield sites.
  • Because the reward function is modular, social factors (local support, housing, commute, environmental impact) could be added as penalty terms, turning the pipeline into a full project-planning tool—though the paper lists these as future work.
  • A decisive test of the method would be to run IfoScout on the actual candidate sites of planned third-generation detectors, with real geological and seismic data, and compare the resulting configurations against the official designs; this would show whether the co-design gains survive production-level constraints.
  • The paper's own caveat about the non-separability of position-dependent noise suggests that a fully coupled optimization—varying site and optics simultaneously—might find configurations the sequential pipeline misses; building a differentiable surrogate of the geometry software would be the natural next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces IfoScout, a two-stage machine-learning pipeline for co-designing ground-based gravitational-wave detector networks. Stage 1 (GWymnasium) uses PPO reinforcement learning on raster terrain data to place two L-shaped interferometers at two Spanish sites, with a reward favoring long arms and penalizing road crossings, tunnels, bridges, exclusion perimeters, and non-optimal relative orientation. Stage 2 (AutoGrav) is a PyTorch reimplementation of PyGWINC and the inspiral-range SenseMonitor algorithm; starting from Advanced LIGO design parameters, it uses projected gradient descent to optimize the ITM/ETM curvature radii and arm half-length toward cavity stability (target g-factor 0.1) and maximum BNS range, constrained to the arm length selected by RL. The authors report three example two-detector networks (C1/C2/C3) with arm lengths 15–44 km, tunnel/support fractions, and BNS ranges up to 1160 Mpc, and argue that this speeds up site surveys. The paper is explicit that it is a proof of concept, with earth curvature, position-dependent Newtonian noise, and geological/geotechnical effects left out.

Significance. If the surrogate and pipeline were validated, IfoScout would be a useful early-design tool for third-generation detectors: it automates a search over site layout and cavity optics, is fast (RL about one hour, DP about one minute per configuration), and the code is publicly released. The division of labor between RL for geography and DP for internal optics is a sensible strategy, and the authors are candid about their simplifications. However, the current quantitative results—Table I and Fig. 3—rest on AutoGrav, which is not validated against PyGWINC, and the sensitivity optimization ignores the actual site’s Newtonian noise. Therefore the central claim is promising but not yet established; the paper reads as a proof-of-concept that needs a validation layer before the specific numbers can be trusted.

major comments (3)
  1. [Differential programming: AutoGrav; Results, Table I and Fig. 3] The reported BNS ranges are produced solely by AutoGrav, a PyTorch port of PyGWINC and inspiral range. The manuscript does not show a single comparison of AutoGrav against PyGWINC or inspiral range, nor a finite-difference gradient check. Since the DP stage uses d(range)/d(parameters), any error in the noise models or in the gradient projection can move the optimizer to a different optimum or inflate the range. A concrete fix: reproduce the Advanced LIGO sensitivity curve and BNS range with AutoGrav and PyGWINC at the initial and optimized parameter sets, report absolute and relative differences, and verify gradients by finite differences over the explored parameter domain. Without this, the 449–1160 Mpc values and the claim to “optimize scientific goals” are not supported.
  2. [Results and Discussion & Conclusions] The DP stage assumes Hanford-site Newtonian noise and Advanced LIGO design parameters, not the Spanish sites; the Discussion then admits that Newtonian and anthropogenic noises are highly position-dependent and that this non-separability “compromises the quality of solutions of sequential optimizations.” Thus, the configurations are not actually optimized for the chosen sites’ scientific environment—they are optimized for Hanford noise with constraints from Spanish terrain. This is not fatal for a proof-of-concept, but the abstract and conclusions should scale back the claim, or the paper should include a sensitivity study showing that the results are robust to replacing Hanford noise with representative site-noise models.
  3. [Reinforcement learning: GWymnasium; Results; Supplemental Material] The RL reward uses arm length as the main science proxy, and the DP stage is then constrained to the RL-selected length. This is a sequential two-step optimization, not joint co-design; in particular, the DP stage cannot choose a more sensitive length that the RL reward missed because the RL reward does not incorporate the BNS range. The final configurations are also manually adjusted (up to 500 m) and the river exclusion threshold for C3-S2 is changed after visual inspection. The paper would be strengthened by an ablation or multi-start test showing that the RL reward correlates with the final BNS range, and by a clear statement of how much of the final design is attributable to the automated pipeline versus human post-processing.
minor comments (5)
  1. [Eq. (1)] The gradient projection formula is under-specified: α(g) is defined only at the endpoints, and the relation between g, g_ini, g_lim, and the distance to the boundary is not stated. Please define the interpolation and how it is updated during optimization.
  2. [Eq. (2)] The notation for the combined g-factor is confusing: the text writes g2 on the left while also defining individual g1 and g2 on the right. Use a distinct symbol such as g^2 or g_prod to avoid the subscript/superscript collision.
  3. [Reinforcement learning: GWymnasium] The network-orientation reward uses a relative orientation of 45°, but the paper only states that using the relative north instead of the great-circle direction introduces a 1.68° deviation. Please give a sentence explaining how this deviation was estimated and why it is acceptable for the reward.
  4. [Results] The sensitivity comparison in Fig. 3 is only against Advanced LIGO and A+. Since the AutoGrav design inherits the Advanced LIGO high-frequency design choices, this is expected, but a comparison against a simple scaled-length Advanced LIGO configuration would more clearly isolate the benefit of the DP stage.
  5. [Supplemental Material] The optimization timescales paragraph mixes training steps, environments, and wall-clock times. Clarify the number of PPO updates per environment and whether the reported times include the generation of terrain features.

Circularity Check

0 steps flagged

No significant circularity: the reported configurations are optimization outputs, not self-referential predictions.

full rationale

The paper's derivation chain contains no step in which a reported result is equivalent to its input by construction. The RL stage optimizes detector layout using a reward whose main scientific proxy is arm length; the DP stage then optimizes mirror radii and arm length to maximize the BNS range as computed by AutoGrav, subject to the RL-selected length and a target cavity g-factor. The BNS ranges in Table I and the sensitivity curves in Fig. 3 are model evaluations of the chosen objective, not parameters fitted to the data they purport to predict. The reader's intuition about a 'circular-ish pipeline' is better described as a coupling limitation: the paper itself notes that newtonian and anthropogenic noise are position-dependent so that sequential optimization is compromised, and that a production-ready pipeline would need a simultaneous treatment. That is a validity and robustness caveat, not a definitional circularity. The self-citations to Refs. [40,41] motivate the co-design paradigm and support the non-separability caveat, but the central numerical results do not reduce to those citations. The main weakness—lack of direct validation of AutoGrav against PyGWINC/inspiral range—is a correctness risk, not a circularity. The paper is a proof-of-concept demonstration whose outputs are generated by an explicit, externally specified optimization pipeline.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 3 invented entities

The central claim rests on novel software artifacts (IfoScout/AutoGrav/GWymnasium) whose behavior depends on several hand-chosen hyperparameters (g-factor target, learning rate, reward weights, terrain thresholds) and on simplified domain assumptions (BNS range as utility, Hanford noise proxy, elevation-only terrain). These are mostly disclosed in the paper, and the paper explicitly frames itself as a proof-of-concept, but the numerical results (Table I, Fig. 3) are not robust to these choices in a demonstrated way.

free parameters (6)
  • target g-factor = 0.1
    The paper sets a target cavity g-factor of 0.1 for the DP optimization, which determines the cavity shape. This is a designer-chosen hyperparameter, not derived from first principles.
  • gradient clipping distance = 50 m before L
    The gradient of arm length is clipped 50 m before the target, a hand-chosen tolerance that affects the final configuration.
  • learning rate = 2e-5
    Constant learning rate for the AutoGrav optimizations; affects the convergence trajectories and thus the reported ranges.
  • RL reward weights = not reported
    The relative weights/penalties for length, tunnels, supports, edge distance and network orientation are encoded in the reward function but not listed as numerical values, so the balance between scientific and engineering terms is not reproducible from the text alone.
  • vertical margin for light earthworks = 5 m
    The terrain model classifies sections as underground/elevated based on a 5 m vertical margin relative to the surface; this threshold drives the reported tunnel/support fractions.
  • river exclusion threshold for C3 site 2 = lowered to 5 m below surface
    The paper explicitly reports lowering the river exclusion threshold for this site after visual inspection of the canal, a post-hoc adjustment that affects whether C3 is considered valid.
axioms (4)
  • domain assumption RL step count (2000 steps) and network training (2e6 steps) are sufficient for a representative search
    The paper trains for 2e6 steps and tests with 2000-step episodes, implicitly assuming this explores the configuration space well enough to identify the best families.
  • domain assumption BNS range (with SNR > 8) is a sufficient utility proxy for scientific value
    The paper states it is 'only interested in maximization, not in obtaining a specific range', treating BNS range as a proxy that ignores sky localization, stochastic backgrounds (except a heuristic orientation term) and other science cases.
  • domain assumption Terrain elevation data from IGN/Google Earth adequately represent construction costs
    Tunnels and supported sections are derived from 2D elevation profiles; the paper acknowledges that geological data and curvature of the earth would be needed for production, and that newtonian/anthropogenic noise couples position to sensitivity.
  • domain assumption LIGO Hanford newtonian noise and Advanced LIGO design are valid starting points for the two Spanish sites
    The paper assumes 'the Hanford site (for newtonian noise) and the original Advanced LIGO design' as a stand-in, which the discussion flags as not production-ready.
invented entities (3)
  • IfoScout pipeline independent evidence
    purpose: Two-stage optimization methodology for detector network co-design
    It is a software artifact with a public repository; its falsifiable handle is that it can be run on other sites and its outputs compared to real designs.
  • AutoGrav independent evidence
    purpose: PyTorch translation of PyGWINC and inspiral range to enable gradient computation
    It is a code artifact; a competent group could validate it against PyGWINC outputs. The paper does not provide a direct numerical validation against PyGWINC in the text, but code is available.
  • GWymnasium independent evidence
    purpose: RL environment simulating interferometer placement on terrain
    It is a software artifact with a public repository; its behavior can be tested by running the provided example.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Co-design of ground-based gravitational wave detector networks." pith.science (2026). https://pith.science/paper/5GOKDQRJ

@misc{pith2026260727254,
  author       = {Pith},
  title        = {Pith review of: Co-design of ground-based gravitational wave detector networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5GOKDQRJ}},
  note         = {Machine review of arXiv:2607.27254}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Discussions around the design philosophy and location of the next generation of ground-based gravitational wave detectors are still underway. In this context, we propose IfoScout, an innovative methodology for detector co-design based on state-of-the-art machine-learning (ML) techniques. We present a two-stage simulation of a network of fictional L-shaped interferometers whose sensitivity is optimized within physical and geographical constraints, indirectly resulting in reducing the costs. To achieve this, we gather publicly available data for two token locations and establish the length and orientation with reinforcement learning (RL). Next, we optimize the internal detector parameters related to cavity stability to achieve the best possible sensitivity by means of differential programming (DP). We make the case that IfoScout could have a positive impact on the final design of new generation detectors (e.g. the Einstein Telescope, the Cosmic Explorer, etc.), given precise data (e.g. geographical and geological maps of chosen sites) and detailed, realistic simulations of the interferometers.

Figures

Figures reproduced from arXiv: 2607.27254 by Daniel Lanchares, Luigi Toffolatti, Lysiane Mornas, Pietro Vischia.

Figure 1
Figure 1. Figure 1: FIG. 1: Diagram of the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Sensitivity curves of the discussed detectors com [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Optimization diagnostics of the various configurations. All optimizations were run with a constant learning [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Summary plot of Configuration 1. Populated areas discussed on the text appear color-coded. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Summary plot of Configuration 3. Note the canal near the 2000 meter mark. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 10 canonical work pages

  1. [1]

    B. P. Abbott et al., Phys. Rev. Lett.116, 061102 (2016), URLhttps://link.aps.org/doi/10.1103/ PhysRevLett.116.061102

    2016

  2. [2]

    Ram´ on y Cajal

    prior to optimization. The radius of the Input Test Mass (ITM) and End Test Mass (ETM) mirrors would be adjusted by the optimization to maintain stability until the desired length is achieved. In this demonstration, we will limit our optimization to these three parameters. g2 :=g 1g2(1−g 1g2)≥0, ( g1 := 1−L/R IT M g2 := 1−L/R ET M, (2) 3 0.0 5.0 10.0 15.0...

    2000

  3. [3]

    B. P. Abbott et al., Phys. Rev. Lett.119, 161101 (2017), URLhttps://link.aps.org/doi/10.1103/ PhysRevLett.119.161101

    2017

  4. [7]

    A. G. Abac et al., The Astrophysical Journal Let- ters995, L18 (2025), URLhttps://doi.org/10.3847/ 2041-8213/ae0c06

    2025

  5. [8]

    A. G. Abac et al., arXiv e-prints (2026), 2605.27223, URL https://arxiv.org/abs/2605.27223

    Pith/arXiv arXiv 2026

  6. [9]

    Aasi et al., Classical and Quantum Gravity32, 074001 (2015), URLhttps://doi.org/10.1088/0264-9381/32/ 7/074001

    J. Aasi et al., Classical and Quantum Gravity32, 074001 (2015), URLhttps://doi.org/10.1088/0264-9381/32/ 7/074001

    work page doi:10.1088/0264-9381/32/ 2015

  7. [10]

    Acernese et al., Classical and Quantum Gravity 32, 024001 (2015), URLhttps://dx.doi.org/10.1088/ 0264-9381/32/2/024001

    F. Acernese et al., Classical and Quantum Gravity 32, 024001 (2015), URLhttps://dx.doi.org/10.1088/ 0264-9381/32/2/024001

    2015

  8. [11]

    Akutsu et al., Nat

    T. Akutsu et al., Nat. Astron.3, 35 (2019), URLhttps: //doi.org/10.1038/s41550-018-0658-y

    doi:10.1038/s41550-018-0658-y 2019

  9. [12]

    B. P. Abbott et al., Annalen der Physik529, 1600209 (2016), URLhttp://dx.doi.org/10.1002/ andp.201600209

    2016

  10. [13]

    B. P. Abbott et al., Physical Review Letters119(2017), URLhttp://dx.doi.org/10.1103/PhysRevLett.119. 141101

    doi:10.1103/physrevlett.119 2017

  11. [14]

    B. P. Abbott et al., Phys. Rev. X9(2019), URLhttp: //dx.doi.org/10.1103/PhysRevX.9.011001

    doi:10.1103/physrevx.9.011001 2019

  12. [15]

    B. P. Abbott et al., The Astrophysical Journal Let- ters848, L12 (2017), URLhttps://doi.org/10.3847/ 2041-8213/aa91c9

    2017

  13. [16]

    A. G. Abac et al., The Astrophysical Journal Letters993, L25 (2025), ISSN 2041-8213, URLhttp://dx.doi.org/ 10.3847/2041-8213/ae0c9c

    doi:10.3847/2041-8213/ae0c9c 2025

  14. [17]

    A. G. Abac et al., The Astrophysical Journal Letters993, L21 (2025), ISSN 2041-8213, URLhttp://dx.doi.org/ 10.3847/2041-8213/ae0d54

    doi:10.3847/2041-8213/ae0d54 2025

  15. [18]

    Abac et al., Physical Review Letters135(2025), ISSN 1079-7114, URLhttp://dx.doi.org/10.1103/ kw5g-d732

    A. Abac et al., Physical Review Letters135(2025), ISSN 1079-7114, URLhttp://dx.doi.org/10.1103/ kw5g-d732

    2025

  16. [19]

    Punturo et al., Classical and Quantum Gravity 27, 194002 (2010), URLhttps://doi.org/10.1088/ 0264-9381/27/19/194002

    M. Punturo et al., Classical and Quantum Gravity 27, 194002 (2010), URLhttps://doi.org/10.1088/ 0264-9381/27/19/194002

    2010

  17. [20]

    E. D. Hall , Galaxies10(2022), ISSN 2075-4434, URL https://www.mdpi.com/2075-4434/10/4/90

    2022

  18. [21]

    Harms et al., The Astrophysical Journal910, 1 (2021), URLhttps://doi.org/10.3847/1538-4357/abe5a7

    J. Harms et al., The Astrophysical Journal910, 1 (2021), URLhttps://doi.org/10.3847/1538-4357/abe5a7

    doi:10.3847/1538-4357/abe5a7 2021

  19. [22]

    Jani et al., arXiv e-prints (2025), 2508.11631, URL https://arxiv.org/abs/2508.11631

    K. Jani et al., arXiv e-prints (2025), 2508.11631, URL https://arxiv.org/abs/2508.11631

    Pith/arXiv arXiv 2025

  20. [23]

    Zhang, C

    X. Zhang, C. Yu, H. Li, S. Kazempour, M. Li, and S. Sun, npj Space Exploration2, 21 (2026), ISSN 3059-3700, URLhttps://doi.org/10.1038/s44453-026-00037-w

    work page doi:10.1038/s44453-026-00037-w 2026

  21. [24]

    Amaro-Seoane et al., arXiv e-prints (2017), 1702.00786, URLhttps://arxiv.org/abs/1702.00786

    P. Amaro-Seoane et al., arXiv e-prints (2017), 1702.00786, URLhttps://arxiv.org/abs/1702.00786

    Pith/arXiv arXiv 2017

  22. [25]

    Luo et al., Classical and Quantum Gravity33, 035010 (2016), URLhttps://doi.org/10.1088/0264-9381/33/ 3/035010

    J. Luo et al., Classical and Quantum Gravity33, 035010 (2016), URLhttps://doi.org/10.1088/0264-9381/33/ 3/035010

    work page doi:10.1088/0264-9381/33/ 2016

  23. [26]

    Sato et al., Journal of Physics: Conference Series 154, 012040 (2009), URLhttps://doi.org/10.1088/ 1742-6596/154/1/012040

    S. Sato et al., Journal of Physics: Conference Series 154, 012040 (2009), URLhttps://doi.org/10.1088/ 1742-6596/154/1/012040

    2009

  24. [27]

    H. Miao, H. Yang, R. X Adhikari, and Y. Chen, Classical and Quantum Gravity31, 165010 (2014), URLhttps: //doi.org/10.1088/0264-9381/31/16/165010

    work page doi:10.1088/0264-9381/31/16/165010 2014

  25. [28]

    Martynov et al., Phys

    D. Martynov et al., Phys. Rev. D99, 102004 (2019), URLhttps://link.aps.org/doi/10.1103/PhysRevD. 99.102004

    doi:10.1103/physrevd 2019

  26. [29]

    Iacovelli, L

    F. Iacovelli, L. Reali, E. Berti, A. Corsi, B. S. Sathyaprakash, and D. Wadekar, arXiv e-prints (2026), 2604.11871, URLhttps://arxiv.org/abs/2604.11871

    Pith/arXiv arXiv 2026

  27. [30]

    Abac et al., Journal of Cosmology and Astroparticle Physics2026, 081 (2026), URLhttps://doi.org/10

    A. Abac et al., Journal of Cosmology and Astroparticle Physics2026, 081 (2026), URLhttps://doi.org/10. 1088/1475-7516/2026/03/081

    2026

  28. [31]

    Santoliquido, J

    F. Santoliquido, J. Tissino, U. Dupletsa, M. Branchesi, and J. Harms, Astronomy & Astrophysics708, A175 (2026), ISSN 1432-0746, URLhttp://dx.doi.org/10. 1051/0004-6361/202558828

    2026

  29. [32]

    Amann et al., Review of Scientific Instruments91, 7 094504 (2020), ISSN 0034-6748, URLhttps://doi.org/ 10.1063/5.0018414

    F. Amann et al., Review of Scientific Instruments91, 7 094504 (2020), ISSN 0034-6748, URLhttps://doi.org/ 10.1063/5.0018414

    work page doi:10.1063/5.0018414 2020

  30. [33]

    Burchartz et al., Bulletin of Engineering Geology and the Environment84, 609 (2025), ISSN 1435-9537, URL https://doi.org/10.1007/s10064-025-04623-2

    R. Burchartz et al., Bulletin of Engineering Geology and the Environment84, 609 (2025), ISSN 1435-9537, URL https://doi.org/10.1007/s10064-025-04623-2

    work page doi:10.1007/s10064-025-04623-2 2025

  31. [34]

    Bristol, C

    W. Bristol, C. Lukinbeal, J. Hoover, K. Daniel, P. Sledge, J. Russell, M. Evans, and on behalf of the Cosmic Ex- plorer Project, Journal of Physics: Conference Series 3177, 012092 (2026), URLhttps://doi.org/10.1088/ 1742-6596/3177/1/012092

    2026

  32. [35]

    Datrier, G

    L. Datrier, G. Lovelace, J. R. Smith, A. Saenz, A. Romero, and for the Cosmic Explorer Project, Journal of Physics: Conference Series3177, 012093 (2026), URL https://doi.org/10.1088/1742-6596/3177/1/012093

    work page doi:10.1088/1742-6596/3177/1/012093 2026

  33. [36]

    Amann, F

    F. Amann, F. Badaracco, R. DeSalvo, L. Naticchioni, A. Paoli, L. Paoli, P. Ruggi, and S. Selleri, Applied Sci- ences12(2022), ISSN 2076-3417, URLhttps://www. mdpi.com/2076-3417/12/17/8827

    2022

  34. [38]

    J. C. Driggers, J. Harms, and R. X. Adhikari, Phys. Rev. D86, 102001 (2012), URLhttps://link.aps.org/doi/ 10.1103/PhysRevD.86.102001

    doi:10.1103/physrevd.86.102001 2012

  35. [39]

    Harms, Living Reviews in Relativity22, 6 (2019), ISSN 1433-8351, URLhttps://doi.org/10

    J. Harms, Living Reviews in Relativity22, 6 (2019), ISSN 1433-8351, URLhttps://doi.org/10. 1007/s41114-019-0022-2

    2019

  36. [40]

    Schillings, S

    P. Schillings, S. Yao, J. Erdmann, and A. Rietbrock, arXiv e-prints (2026), 2603.15424, URLhttps://arxiv. org/abs/2603.15424

    arXiv 2026

  37. [41]

    Dorigo, A

    T. Dorigo, A. Giammanco, P. Vischia, et al., Re- views in Physics10, 100085 (2023), ISSN 2405- 4283, URLhttps://www.sciencedirect.com/science/ article/pii/S2405428323000047

    2023

  38. [42]

    Dorigo, P

    T. Dorigo, P. Vischia, et al. (MODE), submitted to Re- ports on Progress in Physics (2026), 2603.26613

    arXiv 2026

  39. [43]

    Krenn, Y

    M. Krenn, Y. Drori, and R. X. Adhikari, Phys. Rev. X15, 021012 (2025), URLhttps://link.aps.org/doi/ 10.1103/PhysRevX.15.021012

    doi:10.1103/physrevx.15.021012 2025

  40. [44]

    T. D. P. Edwards, K. W. K. Wong, K. K. H. Lam, A. Coogan, D. Foreman-Mackey, M. Isi, and A. Zimmer- man, Phys. Rev. D110, 064028 (2024), URLhttps: //link.aps.org/doi/10.1103/PhysRevD.110.064028

    doi:10.1103/physrevd.110.064028 2024

  41. [45]

    Buchli et al., Science389, 1012–1015 (2025), ISSN 1095-9203, URLhttp://dx.doi.org/10.1126/science

    J. Buchli et al., Science389, 1012–1015 (2025), ISSN 1095-9203, URLhttp://dx.doi.org/10.1126/science. adw1291

    doi:10.1126/science 2025

  42. [46]

    Blondel and V

    M. Blondel and V. Roulet, arXiv e-prints (2025), 2403.14606, URLhttps://arxiv.org/abs/2403.14606

    Pith/arXiv arXiv 2025

  43. [47]

    V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller, arXiv e- prints (2013), 1312.5602, URLhttps://arxiv.org/abs/ 1312.5602

    Pith/arXiv arXiv 2013

  44. [48]

    A. Paszke et al., inProceedings of the 33rd Inter- national Conference on Neural Information Process- ing Systems(Curran Associates Inc., Red Hook, NY, USA, 2019), URLhttps://dl.acm.org/doi/10.5555/ 3454287.3455008

    arXiv 2019

  45. [49]

    LIGO Scientific Collaboration,GWINC (Gravitational Wave Interferometer Noise Calculator)(2020), URL https://gwinc.docs.ligo.org/pygwinc/

    2020

  46. [50]

    Raffin, A

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernes- tus, and N. Dormann, Journal of Machine Learning Re- search22, 1 (2021), URLhttps://jmlr.org/papers/ v22/20-1364.html

    2021

  47. [51]

    Towers et al., arXiv e-prints (2025), 2407.17032, URL https://arxiv.org/abs/2407.17032

    M. Towers et al., arXiv e-prints (2025), 2407.17032, URL https://arxiv.org/abs/2407.17032

    Pith/arXiv arXiv 2025

  48. [52]

    Schlaufman and P

    K. Schlaufman and P. Sutton,Sensemonitor (2004), [Online; accessed 14-Jul-2026], URL https://labcit.ligo.caltech.edu/~jzweizig/dmt/ Monitors/SenseMonitor/index.html

    2004

  49. [53]

    Fairhurst, Classical and Quantum Gravity28, 105021 (2011), URLhttps://doi.org/10.1088/0264-9381/28/ 10/105021

    S. Fairhurst, Classical and Quantum Gravity28, 105021 (2011), URLhttps://doi.org/10.1088/0264-9381/28/ 10/105021

    doi:10.1088/0264-9381/28/ 2011

  50. [54]

    M. J. Box, The Computer Journal8, 42 (1965), ISSN 0010-4620, URLhttps://doi.org/10.1093/comjnl/8. 1.42

    work page doi:10.1093/comjnl/8 1965

  51. [55]

    Ackley and others., Publications of the Astronomical Society of Australia37, e047 (2020), URLhttps://doi

    K. Ackley and others., Publications of the Astronomical Society of Australia37, e047 (2020), URLhttps://doi. org/10.1017/pasa.2020.39

    doi:10.1017/pasa.2020.39 2020

  52. [56]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, arXiv e-prints (2017), 1707.06347, URL https://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  53. [57]

    Iacovelli, E

    F. Iacovelli, E. Belgacem, M. Maggiore, M. Mancarella, and N. Muttoni, Journal of Cosmology and Astroparticle Physics2024, 085 (2024), URLhttps://doi.org/10. 1088/1475-7516/2024/10/085

    2024

  54. [58]

    Branchesi et al., Journal of Cosmology and Astropar- ticle Physics2023, 068 (2023), URLhttps://doi.org/ 10.1088/1475-7516/2023/07/068

    M. Branchesi et al., Journal of Cosmology and Astropar- ticle Physics2023, 068 (2023), URLhttps://doi.org/ 10.1088/1475-7516/2023/07/068

    doi:10.1088/1475-7516/2023/07/068 2023

  55. [59]

    Barsotti, L

    L. Barsotti, L. McCuller, M. Evans, and P. Fritschel,The A+ design curve(2018), [Online; accessed 14-Jul-2026], URLhttps://dcc.ligo.org/LIGO-T1800042/public

    2018

  56. [60]

    Ganapathy et al., Phys

    D. Ganapathy et al., Phys. Rev. X13, 041021 (2023), URLhttps://link.aps.org/doi/10.1103/PhysRevX. 13.041021

    doi:10.1103/physrevx 2023

  57. [61]

    Iacovelli, M

    F. Iacovelli, M. Mancarella, S. Foffa, and M. Maggiore, The Astrophysical Journal Supplement Series263, 2 (2022), ISSN 1538-4365, URLhttp://dx.doi.org/10. 3847/1538-4365/ac9129

    2022

  58. [62]

    Dupletsa, J

    U. Dupletsa, J. Harms, B. Banerjee, M. Branchesi, B. Goncharov, A. Maselli, A. C. S. Oliveira, S. Ronchini, and J. Tissino, Astronomy and Computing42, 100671 (2023), 2205.02499, URLhttps://www.sciencedirect. com/science/article/pii/S2213133722000853

    Pith/arXiv arXiv 2023

  59. [63]

    Begnoni, S

    A. Begnoni, S. Anselmi, M. Pieroni, A. Renzi, and A. Ricciardone, Journal of Cosmology and Astroparti- cle Physics2026, 079 (2026), URLhttps://doi.org/ 10.1088/1475-7516/2026/05/079

    work page doi:10.1088/1475-7516/2026/05/079 2026

  60. [64]

    J. M. S. de Souza and M. Quartin, Journal of Cos- mology and Astroparticle Physics2026, 101 (2026), ISSN 1475-7516, URLhttp://dx.doi.org/10.1088/ 1475-7516/2026/05/101

    2026

  61. [65]

    K. J. Daniel et al., Review of Scientific Instruments96, 014502 (2025), ISSN 0034-6748, URLhttps://doi.org/ 10.1063/5.0242016

    work page doi:10.1063/5.0242016 2025

  62. [66]

    Instituto Geogr´ afico Nacional (IGN),MDT05- cob1, [Online; accessed 14-Jul-2026], URLhttps: //centrodedescargas.cnig.es/CentroDescargas/ modelo-digital-terreno-mdt05-primera-cobertura

    2026

  63. [67]

    Google LLC,Google Earth Pro(2026), URLhttps:// www.google.com/earth/

    2026

  64. [68]

    Lanchares,IfoScout(2026), commit of July 24th, URL https://github.com/Daniel-Lanchares/IfoScout

    D. Lanchares,IfoScout(2026), commit of July 24th, URL https://github.com/Daniel-Lanchares/IfoScout. 8 SUPPLEMENT AL MA TERIAL Optimization timescales The design phase of projects of this magnitude is often long enough to render design optimization times almost negligible. However, our highly parallelizable approach is particularly fast, with the RL model ...

    2026

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.