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

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

Problem Statement
Ground-based gravitational-wave (GW) detectors, such as Advanced LIGO, are susceptible to non-Gaussian transient noise artifacts known as "glitches." These artifacts can mimic astrophysical signals, reduce detector sensitivity, and complicate parameter estimation. While the Gravity Spy project has established a standard for glitch classification using supervised machine learning and citizen science, these approaches are inherently limited to recognizing morphologies they have been explicitly trained on. They cannot identify previously uncharacterized glitch types. The central question addressed in this work is whether the LIGO O4a observing run introduced glitch morphologies not present in earlier runs, specifically those that would be invisible to supervised classifiers trained on historical data.

Methodology
The authors present gravi-signal-ml, a fully unsupervised, zero-shot pipeline designed to characterize instrumental noise transients without labeled training data, domain-specific fine-tuning, or GPU acceleration. The methodology proceeds as follows:

1. Data Preprocessing: The pipeline analyzes 1,277 hours of public strain data from the Hanford (H1) and Livingston (L1) detectors across four independent temporal windows in 2023. Raw data is segmented into 32-second chunks, whitened, bandpass-filtered (20–2000 Hz), and converted into 256×256 pixel Q-transform spectrograms using the cividis colormap to ensure perceptual uniformity under grayscale conversion.
2. Feature Extraction: Spectrograms are encoded using a frozen DINOv2 Vision Transformer (ViT-S/14 with register tokens) pre-trained on natural images. The model is used in evaluation mode with no fine-tuning. The CLS token output is extracted as a 384-dimensional embedding and L2-normalized. The use of register tokens is intended to suppress artifacts common when processing non-natural images.
3. Dimensionality Reduction: Embeddings are projected from 384 to 50 dimensions via Principal Component Analysis (PCA), retaining >98% of variance. A subsequent Uniform Manifold Approximation and Projection (UMAP) reduces the data to 10 dimensions for clustering, using a cosine metric to account for the unit-hypersphere geometry of the embeddings.
4. Clustering: A Dirichlet Process Mixture Model (DPMM) is employed to automatically determine the number of morphological groups. Unlike density-based methods (e.g., HDBSCAN), DPMM avoids sensitivity to minimum cluster size by assigning negligible weight to superfluous components.
5. Validation and Cross-Check:
   • Morphological Cross-Check: Anomalous clusters (those with low log-likelihood scores) are compared against an in-domain reference index built from the Gravity Spy O3b labeled dataset, processed through the exact same pipeline to eliminate domain gaps. Samples are classified as Known, Ambiguous, or Novel based on cosine similarity thresholds.
   • Robustness Tests: The pipeline undergoes ablation studies (grayscale conversion, pixel inversion, intensity scaling), stability analysis (hyperparameter perturbation), and time-slide background tests to rule out statistically significant coincidences between H1 and L1.

Key Contributions

1. Zero-Shot Pipeline: The first application of frozen DINOv2 features to GW glitch morphology characterization, requiring no labeled data or GPU resources.
2. Systematic Validation: A multi-layer validation framework applied consistently across four independent O4a windows, including ablation, stability, and time-slide background estimation.
3. Reproducible Baseline: The provision of a fully documented, open-source pipeline and results for zero-shot glitch characterization.
4. Empirical Observation: The identification of a systematic difference in feature robustness between the H1 and L1 detectors.

Results
The pipeline processed over 188,000 spectrograms across the four sessions. The results yielded a null result regarding novel morphologies:

• No Novel Glitches: No morphologically novel glitch candidates were identified. All anomalous clusters mapped to known Gravity Spy classes with a mean top-1 cosine similarity greater than 0.98.
• Cluster Stability: The DPMM identified 10–16 morphological groups per detector per session. Stability analysis confirmed reproducibility, with mean Adjusted Rand Index (ARI) values exceeding 0.83 across perturbed runs.
• Detector Asymmetry: A notable asymmetry was observed in the robustness of the embeddings to grayscale conversion. L1 embeddings showed high robustness (grayscale ARI > 0.90 in most sessions), whereas H1 embeddings exhibited lower and more variable robustness (grayscale ARI ranging from 0.62 to 0.90). This suggests a structural difference in the H1 noise manifold under DINOv2 feature extraction, potentially linked to known non-stationarity in the H1 noise floor.
• Background Exclusion: Time-slide tests showed no statistically significant H1–L1 coincidences (p ≥ 0.1), excluding correlated instrumental artifacts as the cause for the observed populations.
• Benchmarking: When benchmarked against the Gravity Spy reference, the DPMM achieved an ARI of 0.133. The authors note this moderate score reflects the fundamental divergence between human-defined morphological conventions and intrinsic visual similarity in the embedding space, rather than a failure of the method.

Significance and Claims
The paper claims that its null result is scientifically meaningful. It establishes that, at the sensitivity and temporal resolution of this specific pipeline (32-second windows, ViT-S/14 backbone), the O4a glitch population in the analyzed windows is morphologically consistent with the Gravity Spy O3b catalog. This aligns with previous findings that O4a features elevated rates of known glitch classes (e.g., Low Frequency Lines, 1400Ripples) rather than fundamentally new morphologies.

The authors modestly clarify that this result does not prove novel glitches do not exist in O4a; rather, it constrains their prevalence to below the detection threshold of this approach. Limitations acknowledged include the fixed 32-second window (which may miss sub-second transients), the use of a smaller ViT backbone, and the fact that the analyzed windows cover only ~5% of the full O4a dataset. The work serves as a reproducible baseline for future zero-shot characterization efforts, with the pipeline and all data publicly available. Specifically, the open-source pipeline 'gravi-signal-ml' is available on GitHub at https://github.com/lucacirfeta/dante-gravi-signal-ml, with an archived release at DOI 10.5281/zenodo.20121860.

This study addresses the growing need for scalable, unsupervised tools in the era of expanding GW catalogs; for context, in May 2026, the LIGO-Virgo-KAGRA Collaboration released the GWTC-5.0 catalog, reporting 161 new gravitational-wave events, reflecting the increasing data volume that motivates such automated characterization approaches.