pith. sign in

arxiv: 2605.29510 · v1 · pith:CNAYZCWJnew · submitted 2026-05-28 · 🌌 astro-ph.HE · astro-ph.IM

Time-Domain Deep Learning for Pairwise Identification of Strongly Lensed Gravitational-Wave Candidates

Fan Zhang , Qikai Zhang , Qiyuan Yang , Jiaqing Huang , Yong Yuan , Xilong Fan This is my paper

Pith reviewed 2026-06-29 06:14 UTC · model grok-4.3

classification 🌌 astro-ph.HE astro-ph.IM
keywords gravitational wavesstrong gravitational lensingdeep learningtime-domain analysisSiamese networkbinary black holesEinstein TelescopeLIGO
0
0 comments X

The pith

A Siamese 1D residual network classifies strongly lensed gravitational-wave pairs directly from whitened time-domain strain segments.

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

The paper formulates identification of strongly lensed GW event pairs as a binary classification task on two preprocessed strain segments. It introduces PI-ResNet, a physics-inspired Siamese residual network that processes the data without converting to spectrograms. Training and testing use simulated binary-black-hole signals lensed by point-mass and singular-isothermal-sphere models, injected into LIGO and Einstein Telescope noise. The reported accuracies reach 95.60 percent for SIS lenses and 93.80 percent for PM lenses under ET design noise. These results indicate that direct 1D learning can supply an efficient preselection statistic ahead of full Bayesian comparisons as GW catalogs grow.

Core claim

PI-ResNet encodes pairs of whitened strain segments through a shared residual backbone with Squeeze-and-Excitation modules; the embeddings are compared via absolute differences and Hadamard products to decide whether the pair is lensed. On simulated data the network reaches 95.60 percent accuracy for singular-isothermal-sphere lenses and 93.80 percent for point-mass lenses under Einstein Telescope design noise, while retaining 84.03 percent and 78.25 percent accuracy under simulated LIGO H1-L1 Gaussian noise.

What carries the argument

PI-ResNet: a Siamese one-dimensional residual network whose shared backbone with Squeeze-and-Excitation modules encodes two strain segments, after which absolute feature differences and Hadamard products produce the classification score.

If this is right

  • Exhaustive Bayesian pair comparisons can be restricted to a much smaller set of candidates flagged by the network.
  • Avoiding the intermediate time-frequency image step removes one layer of computational cost from the prescreening pipeline.
  • The same architecture can be retrained on signals from different lens models without changing the input representation.
  • The reported gap between ET and LIGO performance points to the value of domain-adaptation techniques before deployment on real detectors.

Where Pith is reading between the lines

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

  • If the accuracy holds on real data, the method could scale to catalogs containing thousands of events without proportional growth in compute.
  • The time-domain formulation may transfer to other pairwise GW tasks such as coincidence searches across detector networks.
  • Extending the training set to include waveforms with higher-order modes or precession would test whether the learned features remain robust.

Load-bearing premise

Performance measured on simulated binary-black-hole signals lensed by point-mass or singular-isothermal-sphere models and injected into idealized detector noise will translate to real detector data containing unknown noise features and more complex lensing.

What would settle it

A large drop in classification accuracy when the trained model is applied to actual LIGO strain data containing real glitches or to signals lensed by mass distributions more complex than point-mass or SIS would falsify the claim of practical utility.

Figures

Figures reproduced from arXiv: 2605.29510 by Fan Zhang, Jiaqing Huang, Qikai Zhang, Qiyuan Yang, Xilong Fan, Yong Yuan.

Figure 1
Figure 1. Figure 1: Workflow used for the simulated lensed-GW pairwise verification task. The pipeline includes waveform and lensing simulation (Stage 1), noise injection and whitening (Stage 2), Siamese 1D feature extraction and pairwise fusion (Stage 3), and performance evaluation (Stage 4). To quantify the signal strength, we use the optimal signal-to-noise ratio (SNR). For a detector response h˜(f) and the corresponding o… view at source ↗
Figure 2
Figure 2. Figure 2: Representative unwhitened time-domain strain segments for the two images of the same PM-lensed GW event. The gray curves denote the noisy detector strains, and the blue curves denote the corresponding pure templates. The two images show similar chirp morphology but different amplitudes and SNRs, reflecting the different magnifications of the lensed images [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Optimal SNR distributions of the simulated un￾lensed and lensed GW signals. The vertical axis is shown on a logarithmic scale. Lensing magnification shifts part of the lensed population toward higher SNR values, producing a broader high-SNR tail. ages. In the subsequent preprocessing step, these strain segments are whitened and then used as paired inputs to the classifier [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 4
Figure 4. Figure 4: End-to-end PI-ResNet architecture. Left: Two candidate whitened strain segments, denoted as L1 and L2, are processed by a shared Siamese backbone. Right: Enlarged view of the ResBlock, which integrates SE modules. The final output is classified based on the fused pairwise representation. phase-coherence-related multiplicative interactions. It is also closely related to the pairwise consistency tests used i… view at source ↗
Figure 5
Figure 5. Figure 5: ROC and PR curves for SIS and PM lens models in the noisy ET setting. Left: ROC curves, with the TPR axis restricted to 0.8–1.0 for clarity. Right: PR curves. should not be attributed to glitches. Instead, it mainly reflects the difficulty of transferring the classifier to a different detector-noise PSD. These results support the use of PI-ResNet as a lightweight preselection model, while also indicating t… view at source ↗
Figure 6
Figure 6. Figure 6: Confusion-matrix summary for pairwise classification across the main evaluation scenarios. Top row: 1D PI-ResNet evaluated on pure and ET-design Gaussian-noise mock data for the SIS and PM lens models. Second row: 2D SEMD baseline evaluated under the corresponding pure and ET-design Gaussian-noise mock configurations. Third row: 1D PI-ResNet evaluated on the LIGO-comparison datasets, including the pure con… view at source ↗
Figure 8
Figure 8. Figure 8: Pairwise association efficiency, measured by the true-positive rate (TPR), as a function of optimal matched-- filter SNR for the noisy ET validation samples. The TPR is computed in logarithmic SNR bins. The trend shows that the association efficiency remains high over most of the sam￾pled range and tends to saturate at higher SNR. Statistical fluctuations in low-occupancy bins (N < 10 samples) appear as bi… view at source ↗
Figure 9
Figure 9. Figure 9: further shows a representative false-positive pair. Although the two signals are not generated from the same lensed source, they show locally similar peak structures within a narrow time-sample region. This ex￾ample illustrates how unrelated pairs can produce short local waveform similarities that lead to spurious pairwise associations. 4.4. Ablation Study To evaluate the contribution of individual compone… view at source ↗
Figure 10
Figure 10. Figure 10: Cross-regime generalization matrix for SIS and PM lens models under noisy ET conditions. Each entry re￾ports the performance obtained when training on one lens model and testing on either the same or the other lens model. To evaluate whether PI-ResNet captures lensing￾consistency features that transfer across different lens models, we perform cross-regime tests between the SIS and PM datasets. Specificall… view at source ↗
read the original abstract

As gravitational wave (GW) catalogs continue to expand, exhaustive Bayesian comparisons of candidate event pairs become increasingly computationally expensive, which motivates the development of fast prescreening methods for strongly lensed GW searches. We formulate lensed-pair identification as a binary verification problem using two preprocessed strain segments. To address this task, we propose Physics-Inspired ResNet (PI-ResNet), a Siamese one-dimensional residual network for pairwise GW candidate classification. Unlike spectrogram-based prescreening approaches, PI-ResNet operates directly on whitened time-domain strain data and avoids an intermediate time--frequency image representation. A shared residual backbone with Squeeze-and-Excitation (SE) modules encodes the two input segments, and the paired embeddings are compared through absolute feature differences and Hadamard-product interactions. We train and evaluate the model using simulated GW signals from binary black hole mergers lensed by point-mass (PM) and singular isothermal sphere (SIS) lenses, injected into simulated LIGO and Einstein Telescope (ET) detector noise. Under ET design noise, PI-ResNet achieves accuracies of $95.60\%$ for SIS lenses and $93.80\%$ for PM lenses, while maintaining $84.03\%$ and $78.25\%$ accuracy under simulated LIGO H1--L1 Gaussian noise. These results suggest that direct learning from 1D strain data provides an efficient and physically motivated preselection statistic for candidate lensed GW pairs, while also indicating the need for detector-domain adaptation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents Physics-Inspired ResNet (PI-ResNet), a Siamese 1D residual network with Squeeze-and-Excitation modules, for binary classification of whether pairs of gravitational-wave strain segments are strongly lensed. The model is trained and tested on simulated binary black hole signals lensed by point-mass and singular isothermal sphere models, injected into stationary Gaussian noise for LIGO and Einstein Telescope detectors. It reports accuracies of 95.60% (SIS) and 93.80% (PM) under ET noise, and 84.03% and 78.25% under LIGO H1-L1 noise, positioning the method as an efficient prescreening tool for lensed GW candidate pairs.

Significance. If the reported performance holds under more realistic conditions, the approach could significantly reduce the computational burden of searching for strongly lensed GW events by providing a fast, direct time-domain prescreening statistic that avoids exhaustive Bayesian parameter estimation for all pairs. The avoidance of time-frequency representations and the use of physics-inspired architecture are positive aspects. The work highlights the potential of deep learning for GW data analysis but is currently limited by its simulation scope.

major comments (2)
  1. [Abstract] Abstract: The central performance claims (95.60% SIS and 93.80% PM under ET design noise; 84.03% and 78.25% under LIGO H1-L1) are presented without error bars, details on the train/validation/test data splits, or information on hyperparameter search, which are necessary to evaluate the reliability of these figures for the prescreening application.
  2. [Abstract] Abstract: The evaluation uses only stationary Gaussian noise and the simplest analytic lens models (PM and SIS); no results are shown for non-stationary noise features, glitches, or more complex lensing (e.g., NFW profiles or substructure), which is load-bearing for the claim that the method supplies a reliable prescreening statistic for real searches, especially given the paper's own note on the need for detector-domain adaptation.
minor comments (2)
  1. The manuscript would benefit from including a table summarizing the accuracies with standard deviations across multiple runs or cross-validation folds.
  2. Clarify the exact preprocessing steps for whitening the strain data and the precise architecture details (e.g., number of residual blocks, SE module placement) in the methods section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the scope and presentation of our results. We provide point-by-point responses below and indicate the revisions planned for the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claims (95.60% SIS and 93.80% PM under ET design noise; 84.03% and 78.25% under LIGO H1-L1) are presented without error bars, details on the train/validation/test data splits, or information on hyperparameter search, which are necessary to evaluate the reliability of these figures for the prescreening application.

    Authors: We agree that the abstract would be strengthened by including these details. The full manuscript already reports the data splits, hyperparameter optimization procedure, and repeated training runs in the Methods section. In the revision we will add a concise summary of the train/validation/test split ratios, note that hyperparameters were selected via grid search with cross-validation, and include error bars (standard deviation across independent runs) on the quoted accuracies. revision: yes

  2. Referee: [Abstract] Abstract: The evaluation uses only stationary Gaussian noise and the simplest analytic lens models (PM and SIS); no results are shown for non-stationary noise features, glitches, or more complex lensing (e.g., NFW profiles or substructure), which is load-bearing for the claim that the method supplies a reliable prescreening statistic for real searches, especially given the paper's own note on the need for detector-domain adaptation.

    Authors: We acknowledge this limitation of the current study. The manuscript already states that the results are obtained under stationary Gaussian noise and simple lens models and explicitly notes the need for detector-domain adaptation. In the revision we will (i) temper the abstract language to describe the reported figures as performance on idealized simulations, (ii) expand the discussion section with an explicit limitations paragraph, and (iii) add forward-looking statements about planned extensions to non-stationary noise, glitches, and more complex lens profiles. Demonstrating performance on those more realistic cases lies outside the scope of the present work, which focuses on establishing a baseline time-domain approach. revision: partial

Circularity Check

0 steps flagged

No circularity; empirical ML performance on independent simulations

full rationale

The paper trains and evaluates PI-ResNet on simulated BBH signals lensed by PM/SIS models and injected into stationary Gaussian noise. Reported accuracies (e.g., 95.60% SIS under ET noise) are standard test-set metrics on held-out simulations generated independently of the model weights. No equations, self-citations, or ansatzes reduce these numbers to the training inputs by construction. The derivation chain consists of standard supervised learning steps with no load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central performance numbers rest on the domain assumption that simulated lensed signals plus idealized noise are representative, plus the standard assumption that a neural network trained on one distribution will generalize. No free parameters are explicitly named in the abstract beyond the usual network weights; no new physical entities are introduced.

axioms (2)
  • domain assumption Simulated binary-black-hole signals lensed by point-mass or singular-isothermal-sphere models and injected into Gaussian detector noise are sufficiently representative of real observations for the purpose of measuring classification accuracy.
    All reported accuracies are obtained exclusively on such simulations.
  • domain assumption A neural network trained to distinguish lensed from unlensed pairs on simulated data will produce a useful prescreening statistic on real data.
    The paper positions the model as a practical preselection tool without additional validation on real events.

pith-pipeline@v0.9.1-grok · 5822 in / 1561 out tokens · 27709 ms · 2026-06-29T06:14:07.943055+00:00 · methodology

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

39 extracted references · 17 canonical work pages · 4 internal anchors

  1. [1]

    2015, Classical and quantum gravity, 32, 074001

    Aasi, J., Abbott, B., Abbott, R., et al. 2015, Classical and quantum gravity, 32, 074001

    2015

  2. [2]

    P., Abbott, R., Abbott, T

    Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2016, Physical review letters, 116, 061102

    2016

  3. [3]

    P., Abbott, R., Abbott, T., et al

    Abbott, B. P., Abbott, R., Abbott, T., et al. 2019, Physical Review X, 9, 031040 14

    2019

  4. [4]

    D., Abraham, S., et al

    Abbott, R., Abbott, T. D., Abraham, S., et al. 2021, The Astrophysical Journal, 923, 14, doi: 10.3847/1538-4357/ac23db

    work page doi:10.3847/1538-4357/ac23db 2021

  5. [5]

    2023, arXiv preprint arXiv:2304.08393

    Abbott, R., Abe, H., Acernese, F., et al. 2023, arXiv preprint arXiv:2304.08393

    work page arXiv 2023

  6. [6]

    D., et al

    Ashton, G., Hübner, M., Lasky, P. D., et al. 2019, Astrophysical Journal Supplement, 241, 27, doi: 10.3847/1538-4365/ab06fc Astropy Collaboration. 2022, Astrophysical Journal, 935, 167, doi: 10.3847/1538-4357/ac7c74

    work page doi:10.3847/1538-4365/ab06fc 2019

  7. [7]

    2017, Physical Review D, 95, 063512

    Baker, T., & Trodden, M. 2017, Physical Review D, 95, 063512

    2017

  8. [8]

    1993, Advances in neural information processing systems, 6

    Bromley, J., Guyon, I., LeCun, Y., Säckinger, E., & Shah, R. 1993, Advances in neural information processing systems, 6

    1993

  9. [9]

    2025, IEEE Transactions on Pattern Analysis and Machine Intelligence, 47, 6531

    Cevher, V. 2025, IEEE Transactions on Pattern Analysis and Machine Intelligence, 47, 6531

    2025

  10. [10]

    2020, Machine Learning: Science and Technology, 2, 011002

    Cuoco, E., Powell, J., Cavaglià, M., et al. 2020, Machine Learning: Science and Technology, 2, 011002

    2020

  11. [11]

    Cutler, C., & Flanagan, E. E. 1994, Phys. Rev. D, 49, 2658, doi: 10.1103/PhysRevD.49.2658

    work page doi:10.1103/physrevd.49.2658 1994

  12. [12]

    2018, Physical Review D, 98, 104029

    Dai, L., Li, S.-S., Zackay, B., Mao, S., & Lu, Y. 2018, Physical Review D, 98, 104029

    2018

  13. [13]

    Deguchi, S., & Watson, W. D. 1986, Astrophysical Journal, Part 1 (ISSN 0004-637X), vol. 307, Aug. 1, 1986, p. 30-37., 307, 30

    1986

  14. [14]

    2005, in Relativistic Gravity Research With Emphasis on Experiments and Observations: Proceedings of the 81 WE-Heraeus-Seminar Held at the Physikzentrum, Bad

    Ehlers, J., & Schneider, P. 2005, in Relativistic Gravity Research With Emphasis on Experiments and Observations: Proceedings of the 81 WE-Heraeus-Seminar Held at the Physikzentrum, Bad

    2005

  15. [15]

    Honnef, Germany 2–6 September 1991, Springer, 1–45

    1991

  16. [16]

    2018, Physical Review D, 97, 044039

    George, D., & Huerta, E. 2018, Physical Review D, 97, 044039

    2018

  17. [17]

    J., & Ajith, P

    Goyal, S., Kapadia, S. J., & Ajith, P. 2021, Physical Review D, 104, 124057

    2021

  18. [18]

    A., Haris, K., Ng, K

    Hannuksela, O. A., Haris, K., Ng, K. K. Y., et al. 2019, Astrophys. J. Lett., 874, L2, doi: 10.3847/2041-8213/ab0c0f

    work page doi:10.3847/2041-8213/ab0c0f 2019

  19. [19]

    Identifying strongly lensed gravitational wave signals from binary black hole mergers

    Ajith, P. 2018, arXiv preprint arXiv:1807.07062

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  20. [20]

    2018, in Proceedings of the IEEE conference on computer vision and pattern recognition, 7132–7141

    Hu, J., Shen, L., & Sun, G. 2018, in Proceedings of the IEEE conference on computer vision and pattern recognition, 7132–7141

    2018

  21. [21]

    A., Haris, K., & Van Den Broeck, C

    Janquart, J., Hannuksela, O. A., Haris, K., & Van Den Broeck, C. 2021, Monthly Notices of the Royal Astronomical Society, 506, 5430

    2021

  22. [22]

    2019, PhRvD, 100, 024059, doi: 10.1103/PhysRevD.100.024059

    Khan, S., Chatziioannou, K., Hannam, M., & Ohme, F. 2019, PhRvD, 100, 024059, doi: 10.1103/PhysRevD.100.024059

    work page doi:10.1103/physrevd.100.024059 2019

  23. [23]

    Kim, K., Lee, J., Yuen, R. S. H., Hannuksela, O. A., & Li, T. G. F. 2021, Astrophys. J., 915, 119, doi: 10.3847/1538-4357/ac0143

    work page doi:10.3847/1538-4357/ac0143 2021

  24. [24]

    2026, The Astrophysical Journal Supplement Series, 283, 31, doi: 10.3847/1538-4365/ae4340

    Li, D., Liu, T., Liu, A., et al. 2026, The Astrophysical Journal Supplement Series, 283, 31, doi: 10.3847/1538-4365/ae4340

    work page doi:10.3847/1538-4365/ae4340 2026

  25. [25]

    2018, MNRAS, 476, 2220, doi: 10.1093/mnras/sty411 38

    Li, S.-S., Mao, S., Zhao, Y., & Lu, Y. 2018, Mon. Not. R. Astron. Soc., 476, 2220, doi: 10.1093/mnras/sty411

    work page doi:10.1093/mnras/sty411 2018

  26. [26]

    2017, Nat

    Liao, K., Fan, X.-L., Ding, X., Biesiada, M., & Zhu, Z.-H. 2017, Nat. Commun., 8, 1148, doi: 10.1038/s41467-017-01152-9

    work page doi:10.1038/s41467-017-01152-9 2017

  27. [27]

    2024, Monthly Notices of the Royal Astronomical Society, 535, 990

    Magare, S., More, A., & Choudhary, S. 2024, Monthly Notices of the Royal Astronomical Society, 535, 990

    2024

  28. [28]

    2017, Physical Review D, 95, 042001

    Messick, C., Blackburn, K., Brady, P., et al. 2017, Physical Review D, 95, 042001

    2017

  29. [29]

    Ohanian, H. C. 1974, International Journal of Theoretical Physics, 9, 425

    1974

  30. [30]

    A., & Li, T

    Pagano, G., Hannuksela, O. A., & Li, T. G. 2020, Astronomy & Astrophysics, 643, A167

    2020

  31. [31]

    Harry, I. W. 2020, Monthly Notices of the Royal Astronomical Society, 495, 3740 Planck Collaboration. 2020, Astronomy & Astrophysics, 641, A6, doi: 10.1051/0004-6361/201833910

    work page doi:10.1051/0004-6361/201833910 2020

  32. [32]

    2021, PhRvD, 103, 104056, doi: 10.1103/PhysRevD.103.104056 Pürrer, M., Hannam, M., & Ohme, F

    Pratten, G., García-Quirós, C., Colleoni, M., et al. 2021, PhRvD, 103, 104056, doi: 10.1103/PhysRevD.103.104056

    work page doi:10.1103/physrevd.103.104056 2021

  33. [33]

    2010, Classical and Quantum Gravity, 27, 194002

    Punturo, M., Abernathy, M., Acernese, F., et al. 2010, Classical and Quantum Gravity, 27, 194002

    2010

  34. [34]

    Cosmic Explorer: The U.S. Contribution to Gravitational-Wave Astronomy beyond LIGO

    Reitze, D., Adhikari, R. X., Ballmer, S., et al. 2019, arXiv preprint arXiv:1907.04833

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  35. [35]

    The GstLAL Search Analysis Methods for Compact Binary Mergers in Advanced LIGO's Second and Advanced Virgo's First Observing Runs

    Sachdev, S., Caudill, S., Fong, H., et al. 2019, arXiv preprint arXiv:1901.08580

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  36. [36]

    2010, Physical review letters, 105, 251101

    Sereno, M., Sesana, A., Bleuler, A., et al. 2010, Physical review letters, 105, 251101

    2010

  37. [37]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Simonyan, K., & Zisserman, A. 2014, CoRR, abs/1409.1556. https://api.semanticscholar.org/CorpusID:14124313

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  38. [38]

    2003, ApJ, 595, 1039, doi: 10.1086/377430

    Takahashi, R., & Nakamura, T. 2003, Astrophys. J., 595, 1039, doi: 10.1086/377430

    work page doi:10.1086/377430 2003

  39. [39]

    A., Nitz, A

    Usman, S. A., Nitz, A. H., Harry, I. W., et al. 2016, Classical and Quantum Gravity, 33, 215004

    2016