Unsupervised Morphological Characterization of Gravitational-Wave Glitches in LIGO O4a Using Frozen DINOv2 Features

Luca Cirfeta

arxiv: 2605.28572 · v1 · pith:BZKQHAUVnew · submitted 2026-05-27 · 🌌 astro-ph.IM

Unsupervised Morphological Characterization of Gravitational-Wave Glitches in LIGO O4a Using Frozen DINOv2 Features

Luca Cirfeta This is my paper

Pith reviewed 2026-06-29 09:54 UTC · model grok-4.3

classification 🌌 astro-ph.IM

keywords gravitational wave glitchesLIGO O4aunsupervised clusteringDINOv2 embeddingsQ-transform spectrogramsmorphology characterizationGravity Spy classes

0 comments

The pith

No morphologically novel glitches appear in LIGO O4a data beyond known classes.

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

The paper tests whether the O4a observing run introduced glitch morphologies absent from earlier LIGO data. It builds an unsupervised pipeline that turns Q-transform spectrograms into 384-dimensional embeddings via a frozen DINOv2 Vision Transformer, reduces the embeddings with PCA and UMAP, and groups them with a Dirichlet Process Mixture Model. Across more than 188,000 spectrograms from 1,277 hours of Hanford and Livingston data, every cluster maps to an existing Gravity Spy class at cosine similarity above 0.98. The result supplies a documented, reproducible baseline for checking future runs for unexpected noise features.

Core claim

Application of frozen DINOv2 embeddings to Q-transform spectrograms of O4a strain data, followed by PCA-UMAP projection and Dirichlet Process Mixture Model clustering, yields no clusters that deviate from known Gravity Spy morphologies; all anomalous clusters achieve cosine similarity greater than 0.98 to reference classes, establishing a null result for novel glitch morphologies in the examined dataset.

What carries the argument

Frozen DINOv2 Vision Transformer (ViTS/14 with register tokens) embeddings from Q-transform spectrograms, used to drive unsupervised PCA-UMAP-DPMM clustering and similarity checks against a Gravity Spy O3b reference index.

If this is right

The pipeline supplies a validated zero-shot baseline that can be reapplied to later observing runs without retraining.
L1 embeddings maintain high cluster stability (ARI above 0.90) across all four sessions examined.
H1 embeddings exhibit lower and more variable stability, pointing to detector-specific structure in the noise manifold.
Time-slide tests show no statistically significant H1-L1 coincidences in any session.

Where Pith is reading between the lines

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

If the embedding space proves complete, any future cluster falling outside known classes would indicate a genuine change in detector behavior rather than an analysis artifact.
The observed difference in robustness between H1 and L1 suggests separate per-detector reference sets may improve sensitivity in follow-up work.
The same embedding-plus-clustering approach could be tested on other transient signals represented as time-frequency images.

Load-bearing premise

The frozen DINOv2 embeddings extracted from the spectrograms already contain enough information to reveal any new glitch morphologies that might exist.

What would settle it

Discovery of at least one cluster whose embeddings have cosine similarity below 0.98 to every known Gravity Spy class would falsify the claim that no novel morphologies are present.

Figures

Figures reproduced from arXiv: 2605.28572 by Luca Cirfeta.

**Figure 1.** Figure 1: Example Q-transform spectrogram (32 s, H1, cividis colormap). The x-axis represents time samples and the y-axis frequency in Hz. This format is used as input to the DINOv2 encoder. Each 256×256 spectrogram is encoded using a frozen DINOv2 Vision Transformer with register tokens (dinov2 vits14 reg, ViT-S/14) (Oquab et al. 2023; Darcet et al. 2024). The model is loaded via torch.hub with frozen weights in e… view at source ↗

**Figure 2.** Figure 2: UMAP 2D projection of L1 embeddings from session 20260524 200219 (21,985 spectrograms, 10 clusters). Each color represents a distinct morphological cluster identified by DPMM. Black crosses (×) indicate samples with log-likelihood below the 1st percentile of the session-wide distribution (57 samples). All clusters show high similarity to known Gravity Spy classes (mean top-1 cosine similarity > 0.98). No … view at source ↗

read the original abstract

A central open question in gravitational-wave detector characterization is whether the O4a observing run has introduced glitch morphologies not present in earlier runs. We present gravi-signal-ml, an open-source pipeline for unsupervised morphological characterization of instrumental noise transients (glitches) in LIGO gravitational-wave data, applied to 1,277 hours of public O4a strain data from the Hanford and Livingston detectors. The pipeline extracts 384-dimensional visual embeddings from Q-transform spectrograms using a frozen DINOv2 Vision Transformer with register tokens (ViTS/14), requiring no labeled training data. Embeddings are projected via PCA and UMAP with cosine metric, then clustered using a Dirichlet Process Mixture Model (DPMM). Cluster robustness is systematically assessed through ablation studies, stability analysis across hyperparameter perturbations, and morphological cross-check against an in-domain Gravity Spy O3b reference index. A time-slide background test excludes statistically significant H1--L1 coincidences ($p \geq 0.1$) in all sessions. Across 188,000+ spectrograms, no morphologically novel glitch candidates were identified -- all anomalous clusters map to known Gravity Spy classes with cosine similarity $> 0.98$. L1 embeddings show consistently high robustness (ablation ARI $> 0.90$ in all four sessions), while H1 exhibits lower and more variable grayscale ablation ARI ($\sim 0.68$--$0.90$), suggesting a structural difference in the H1 noise manifold under DINOv2 feature extraction. This null result, obtained with a fully validated pipeline, establishes a reproducible baseline for zero-shot glitch morphology characterization in O4a data. The pipeline and all results are publicly available at https://github.com/lucacirfeta/dante-gravi-signal-ml DOI: https://doi.org/10.5281/zenodo.20121860.

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.

Desk Editor's Note private letter to a colleague

DINOv2 embeddings on O4a glitches produce a documented null result with good checks, but the pipeline's ability to flag truly novel morphologies remains untested.

read the letter

The main thing to know is that the authors took public O4a strain data, ran Q-transform spectrograms through a frozen DINOv2 ViT, projected with PCA+UMAP, clustered with DPMM, and found every anomalous cluster mapped back to known Gravity Spy classes at cosine similarity above 0.98. No new morphologies appeared across 188k+ examples. The pipeline is fully open, which is the real deliverable.

They do the reproducibility part right. Ablation studies on grayscale inputs, stability runs across hyperparameter tweaks, direct cross-check to the O3b Gravity Spy index, and time-slide background tests that show no significant H1-L1 coincidences. L1 embeddings hold up with ARI above 0.90; H1 is weaker and more variable, which they report plainly. That level of internal checking is better than most transfer-learning applications in this area.

The soft spot is the embedding assumption itself. DINOv2 was trained on natural RGB photos, not time-frequency plots, and the grayscale ablation already drops ARI noticeably on H1. Nothing in the reported tests checks whether a genuinely new glitch pattern could produce near-identical activations and get absorbed into an existing cluster. Without a positive control, such as injected synthetic morphologies outside the Gravity Spy span, the null result is only as strong as the feature space allows. The H1 variability flags the same issue from another angle.

This is for people who maintain glitch catalogs or run data-quality pipelines at LIGO. A reader who needs a reproducible baseline on O4a data will find it useful; someone looking for a new framework or a field-changing claim will not. The work is coherent on its own terms and the checks are honest, so it deserves a serious referee even if the embedding limitation needs to be stated more clearly in revision.

Referee Report

2 major / 1 minor

Summary. The manuscript presents an open-source pipeline (gravi-signal-ml) for unsupervised characterization of LIGO glitches in 1,277 hours of O4a data. It extracts 384-dimensional embeddings from Q-transform spectrograms using a frozen DINOv2 ViT-S/14 model, projects them via PCA+UMAP, and clusters with a Dirichlet Process Mixture Model. Supported by ablation studies, stability analysis, Gravity Spy O3b cross-validation, and time-slide background tests, the central claim is a null result: across 188,000+ spectrograms, all anomalous clusters map to known Gravity Spy classes with cosine similarity >0.98, with no morphologically novel candidates identified. L1 shows high robustness while H1 exhibits lower and more variable performance.

Significance. If the embedding space is shown to be sensitive to potential new morphologies, the work supplies a reproducible, label-free baseline for glitch morphology searches in future runs, strengthened by the fully public code, data, and results at the cited GitHub and Zenodo links. This is a concrete contribution to detector characterization in the gravitational-wave instrumentation community.

major comments (2)

[Abstract and pipeline description] Abstract and pipeline description: The null result (no new morphologies identified) is load-bearing on the assumption that any genuinely novel glitch morphology would produce a distinct DPMM cluster with cosine similarity ≤0.98 to the Gravity Spy O3b reference. The manuscript reports no test (e.g., synthetic injection of time-frequency patterns outside the Gravity Spy span) confirming that frozen DINOv2 embeddings on spectrograms would separate such cases rather than collapsing them into existing clusters.
[Ablation studies] Ablation studies: The grayscale ablation ARI drop for H1 (0.68–0.90) versus L1 (>0.90) demonstrates that the feature extractor is sensitive to input representation, yet no analysis quantifies how this variability impacts the detection threshold for novel morphologies or whether the reported cross-validation remains valid under the lower-ARI regime.

minor comments (1)

The manuscript would benefit from an explicit limitations subsection discussing the domain shift between DINOv2's natural-image pretraining and Q-transform spectrograms.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the scope and limitations of our null result. We address each major comment below and will revise the manuscript accordingly to improve transparency.

read point-by-point responses

Referee: The null result (no new morphologies identified) is load-bearing on the assumption that any genuinely novel glitch morphology would produce a distinct DPMM cluster with cosine similarity ≤0.98 to the Gravity Spy O3b reference. The manuscript reports no test (e.g., synthetic injection of time-frequency patterns outside the Gravity Spy span) confirming that frozen DINOv2 embeddings on spectrograms would separate such cases rather than collapsing them into existing clusters.

Authors: We agree this assumption is central to interpreting the null result. The pipeline's validity for known classes is supported by the Gravity Spy O3b cross-validation (high ARI and cosine similarities >0.98) and the fact that DINOv2 was pretrained on a broad image corpus, enabling zero-shot feature extraction. However, we did not conduct synthetic injections of morphologies outside the Gravity Spy taxonomy. In the revised manuscript we will add an explicit limitations subsection discussing this point, including a qualitative assessment of embedding sensitivity based on existing ablation stability and a recommendation for future controlled-injection studies. This addition will not change the reported O4a findings but will better bound the claim. revision: yes
Referee: The grayscale ablation ARI drop for H1 (0.68–0.90) versus L1 (>0.90) demonstrates that the feature extractor is sensitive to input representation, yet no analysis quantifies how this variability impacts the detection threshold for novel morphologies or whether the reported cross-validation remains valid under the lower-ARI regime.

Authors: The referee is correct that the reported H1 ARI variability indicates greater sensitivity to input representation. We already highlight this difference in the manuscript as evidence of a structural distinction in the H1 noise manifold. We did not, however, propagate the ARI variation into a quantitative estimate of its effect on novel-morphology detection thresholds or re-evaluate cross-validation robustness specifically in the lower-ARI regime. In revision we will extend the ablation section with a short analysis (e.g., stability of cluster assignments under ARI-thresholded subsampling) and a brief statement on how this affects interpretation of the H1 null result. This will be added without altering the core results or conclusions. revision: yes

Circularity Check

0 steps flagged

No circularity: external reference index and time-slide tests provide independent validation

full rationale

The paper's central null result (no novel morphologies across 188k+ spectrograms) is obtained by clustering DINOv2 embeddings with DPMM and then mapping clusters to an external Gravity Spy O3b reference index via cosine similarity >0.98, supplemented by time-slide background tests (p>=0.1) and ablation studies. No step reduces by construction to a fitted parameter, self-definition, or self-citation chain; the pipeline is unsupervised with a frozen pretrained extractor, and the comparison benchmark is independent of the O4a data under analysis. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim depends on the transferability of natural-image DINOv2 features to scientific spectrograms and the completeness of the Gravity Spy reference catalog; these are standard domain assumptions rather than new fitted parameters or invented entities.

axioms (1)

domain assumption Frozen DINOv2 embeddings on Q-transform spectrograms capture sufficient morphological information to detect novel glitch classes if present
Central to the zero-shot pipeline and the interpretation of the null result.

pith-pipeline@v0.9.1-grok · 5880 in / 1338 out tokens · 35675 ms · 2026-06-29T09:54:59.016141+00:00 · methodology

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Patch-Level DINOv2 Scoring for Gravitational-Wave Glitch Detection: Breaking the Signal Dilution Barrier via Vector-Quantized Local Feature Indexing
astro-ph.IM 2026-06 unverdicted novelty 3.0

Patch-level top-k similarity scoring against a vector-quantized DINOv2 reference index yields KS=0.963 separation for extended glitch morphologies on LIGO O4a data, addressing global CLS token dilution.
Sensitivity Limits and Operational Threshold Calibration for DINOv2-based Gravitational-Wave Glitch Characterization: A Strain-Domain Mock Data Challenge on LIGO O4a
astro-ph.IM 2026-06 unverdicted novelty 3.0

Mock data challenge shows DINOv2 pipeline recovers high-SNR anisotropic glitches under dynamic thresholding but yields zero recall for all morphologies under a low-FPR operational threshold due to global average pooling.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages · cited by 2 Pith papers · 3 internal anchors

[1]

Campello, R. J. G. B., Moulavi, D., & Sander, J. 2013, Advances in Knowledge Discovery and Data Mining, Lecture Notes in Computer Science, 7819, 160, https://doi.org/10.1007/978-3-642-37456-2 14

work page doi:10.1007/978-3-642-37456-2 2013
[2]

H., et al

Cabero, M., Lundgren, A., Nitz, A. H., et al. 2019, Classical and Quantum Gravity, 36, 15, https://doi.org/10.1088/1361-6382/ab2e14

work page doi:10.1088/1361-6382/ab2e14 2019
[3]

W., Bahaadini, S., Rohani, N., et al

Coughlin, M. W., Bahaadini, S., Rohani, N., et al. 2019, Physical Review D, 99, 082002, https://doi.org/10.1103/PhysRevD.99.082002

work page doi:10.1103/physrevd.99.082002 2019
[4]

Vision Transformers Need Registers

Darcet, T., Oquab, M., Doup´ e, E., & Bourdoukan, R. 2024, International Conference on Learning Representations (ICLR), arXiv:2309.16588

work page internal anchor Pith review Pith/arXiv arXiv 2024
[5]

2020,Machine Learning: Science and Technology, 2, 015004, https://doi.org/10.1088/2632-2153/abab5f

Essick, R., Godwin, P., Hanna, C., et al. 2020,Machine Learning: Science and Technology, 2, 015004, https://doi.org/10.1088/2632-2153/abab5f

work page doi:10.1088/2632-2153/abab5f 2020
[6]

Ferguson, T. S. 1973,The Annals of Statistics, 1, 209, https://doi.org/10.1214/aos/1176342360

work page doi:10.1214/aos/1176342360 1973
[7]

B., et al

Glanzer, J., Banagiri, S., Coughlin, S. B., et al. 2023, Classical and Quantum Gravity, 40, 065004, https://doi.org/10.1088/1361-6382/acb633

work page doi:10.1088/1361-6382/acb633 2023
[8]

2026,gravi-signal-ml: Unsupervised Morphological Characterization of Gravitational-Wave

Cirfeta, L. 2026,gravi-signal-ml: Unsupervised Morphological Characterization of Gravitational-Wave

2026
[9]

Glitches, Zenodo, https://doi.org/10.5281/zenodo.20121860 LIGO Scientific Collaboration, Virgo Collaboration, & KAGRA Collaboration 2023,Gravitational Wave Open Science Center O4a Data Release, https://gwosc.org

work page doi:10.5281/zenodo.20121860 2023
[10]

Macas, R., Ashton, G., & the gwpy Developers 2021,gwpy: A Python package for gravitational-wave astrophysics, https://doi.org/10.5281/zenodo.5504573

work page doi:10.5281/zenodo.5504573 2021
[11]

2024,arXiv e-prints, arXiv:2404.15552 LIGO Scientific Collaboration 2015,Classical and Quantum Gravity, 32, 074001, https://doi.org/10.1088/0264-9381/32/7/074001

Li, T., et al. 2024,arXiv e-prints, arXiv:2404.15552 LIGO Scientific Collaboration 2015,Classical and Quantum Gravity, 32, 074001, https://doi.org/10.1088/0264-9381/32/7/074001

work page doi:10.1088/0264-9381/32/7/074001 2024
[12]

UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

McInnes, L., Healy, J., & Melville, J. 2018,arXiv e-prints, arXiv:1802.03426 Nu˜ nez, J. R., Anderton, C. R., & Renslow, R. S. 2018, PLOS ONE, 13, e0199239, https://doi.org/10.1371/journal.pone.0199239

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1371/journal.pone.0199239 2018
[13]

DINOv2: Learning Robust Visual Features without Supervision

Oquab, M., Darcet, T., Moutakanni, T., et al. 2023,arXiv e-prints, arXiv:2304.07193 8Cirfeta

work page internal anchor Pith review Pith/arXiv arXiv 2023
[14]

2019,Advances in Neural Information Processing Systems, 32, https://pytorch.org

Paszke, A., Gross, S., Massa, F., et al. 2019,Advances in Neural Information Processing Systems, 32, https://pytorch.org

2019
[15]

2011, Journal of Machine Learning Research, 12, 2825

Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825

2011
[16]

2025,arXiv e-prints, arXiv:2409.02831

Soni, S., et al. 2025,arXiv e-prints, arXiv:2409.02831

work page arXiv 2025
[17]

R., Millman, K

Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020,Nature, 585, 357, https://doi.org/10.1038/s41586-020-2649-2

work page doi:10.1038/s41586-020-2649-2 2020
[18]

2017,Classical and Quantum Gravity, 34, 064003, https://doi.org/10.1088/1361-6382/aa5cea

Zevin, M., Coughlin, S., Bahaadini, S., et al. 2017,Classical and Quantum Gravity, 34, 064003, https://doi.org/10.1088/1361-6382/aa5cea

work page doi:10.1088/1361-6382/aa5cea 2017

[1] [1]

Campello, R. J. G. B., Moulavi, D., & Sander, J. 2013, Advances in Knowledge Discovery and Data Mining, Lecture Notes in Computer Science, 7819, 160, https://doi.org/10.1007/978-3-642-37456-2 14

work page doi:10.1007/978-3-642-37456-2 2013

[2] [2]

H., et al

Cabero, M., Lundgren, A., Nitz, A. H., et al. 2019, Classical and Quantum Gravity, 36, 15, https://doi.org/10.1088/1361-6382/ab2e14

work page doi:10.1088/1361-6382/ab2e14 2019

[3] [3]

W., Bahaadini, S., Rohani, N., et al

Coughlin, M. W., Bahaadini, S., Rohani, N., et al. 2019, Physical Review D, 99, 082002, https://doi.org/10.1103/PhysRevD.99.082002

work page doi:10.1103/physrevd.99.082002 2019

[4] [4]

Vision Transformers Need Registers

Darcet, T., Oquab, M., Doup´ e, E., & Bourdoukan, R. 2024, International Conference on Learning Representations (ICLR), arXiv:2309.16588

work page internal anchor Pith review Pith/arXiv arXiv 2024

[5] [5]

2020,Machine Learning: Science and Technology, 2, 015004, https://doi.org/10.1088/2632-2153/abab5f

Essick, R., Godwin, P., Hanna, C., et al. 2020,Machine Learning: Science and Technology, 2, 015004, https://doi.org/10.1088/2632-2153/abab5f

work page doi:10.1088/2632-2153/abab5f 2020

[6] [6]

Ferguson, T. S. 1973,The Annals of Statistics, 1, 209, https://doi.org/10.1214/aos/1176342360

work page doi:10.1214/aos/1176342360 1973

[7] [7]

B., et al

Glanzer, J., Banagiri, S., Coughlin, S. B., et al. 2023, Classical and Quantum Gravity, 40, 065004, https://doi.org/10.1088/1361-6382/acb633

work page doi:10.1088/1361-6382/acb633 2023

[8] [8]

2026,gravi-signal-ml: Unsupervised Morphological Characterization of Gravitational-Wave

Cirfeta, L. 2026,gravi-signal-ml: Unsupervised Morphological Characterization of Gravitational-Wave

2026

[9] [9]

Glitches, Zenodo, https://doi.org/10.5281/zenodo.20121860 LIGO Scientific Collaboration, Virgo Collaboration, & KAGRA Collaboration 2023,Gravitational Wave Open Science Center O4a Data Release, https://gwosc.org

work page doi:10.5281/zenodo.20121860 2023

[10] [10]

Macas, R., Ashton, G., & the gwpy Developers 2021,gwpy: A Python package for gravitational-wave astrophysics, https://doi.org/10.5281/zenodo.5504573

work page doi:10.5281/zenodo.5504573 2021

[11] [11]

2024,arXiv e-prints, arXiv:2404.15552 LIGO Scientific Collaboration 2015,Classical and Quantum Gravity, 32, 074001, https://doi.org/10.1088/0264-9381/32/7/074001

Li, T., et al. 2024,arXiv e-prints, arXiv:2404.15552 LIGO Scientific Collaboration 2015,Classical and Quantum Gravity, 32, 074001, https://doi.org/10.1088/0264-9381/32/7/074001

work page doi:10.1088/0264-9381/32/7/074001 2024

[12] [12]

UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

McInnes, L., Healy, J., & Melville, J. 2018,arXiv e-prints, arXiv:1802.03426 Nu˜ nez, J. R., Anderton, C. R., & Renslow, R. S. 2018, PLOS ONE, 13, e0199239, https://doi.org/10.1371/journal.pone.0199239

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1371/journal.pone.0199239 2018

[13] [13]

DINOv2: Learning Robust Visual Features without Supervision

Oquab, M., Darcet, T., Moutakanni, T., et al. 2023,arXiv e-prints, arXiv:2304.07193 8Cirfeta

work page internal anchor Pith review Pith/arXiv arXiv 2023

[14] [14]

2019,Advances in Neural Information Processing Systems, 32, https://pytorch.org

Paszke, A., Gross, S., Massa, F., et al. 2019,Advances in Neural Information Processing Systems, 32, https://pytorch.org

2019

[15] [15]

2011, Journal of Machine Learning Research, 12, 2825

Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825

2011

[16] [16]

2025,arXiv e-prints, arXiv:2409.02831

Soni, S., et al. 2025,arXiv e-prints, arXiv:2409.02831

work page arXiv 2025

[17] [17]

R., Millman, K

Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020,Nature, 585, 357, https://doi.org/10.1038/s41586-020-2649-2

work page doi:10.1038/s41586-020-2649-2 2020

[18] [18]

2017,Classical and Quantum Gravity, 34, 064003, https://doi.org/10.1088/1361-6382/aa5cea

Zevin, M., Coughlin, S., Bahaadini, S., et al. 2017,Classical and Quantum Gravity, 34, 064003, https://doi.org/10.1088/1361-6382/aa5cea

work page doi:10.1088/1361-6382/aa5cea 2017