The Sound of Noise: Leveraging the Inductive Bias of Pre-trained Audio Transformers for Glitch Identification in LIGO

TL;DR

This work tackles GW detector glitches and the label bottleneck by recasting strain data as audio and transferring knowledge from a pre-trained Audio Spectrogram Transformer (AST) using LoRA-based PEFT. It processes strain into log-mel spectrograms of size $1024 \times 128$, yielding $D=768$ embeddings that are reduced via PCA and visualized with $t$-SNE for clustering. The results show that the fine-tuned embeddings improve class separability, and injected GW signals occupy regions distinct from glitches, enabling data-efficient anomaly detection. Cross-run tests with Omicron triggers from O4 demonstrate generalization of the learned morphology manifold, supporting a low-latency detector-characterization workflow that mitigates the label bottleneck.

Abstract

Transient noise artifacts, or glitches, fundamentally limit the sensitivity of gravitational-wave (GW) interferometers and can mimic true astrophysical signals, particularly the short-duration intermediate-mass black hole (IMBH) mergers. Current glitch classification methods, such as Gravity Spy, rely on supervised models trained from scratch using labeled datasets. These approaches suffer from a significant ``label bottleneck," requiring massive, expertly annotated datasets to achieve high accuracy and often struggling to generalize to new glitch morphologies or exotic GW signals encountered in observing runs. In this work, we present a novel cross-domain framework that treats GW strain data through the lens of audio processing. We utilize the Audio Spectrogram Transformer (AST), a model pre-trained on large-scale audio datasets, and adapt it to the GW domain. Instead of learning time-frequency features from scratch, our method exploits the strong inductive bias inherent in pre-trained audio models, transferring learned representations of natural sound to the characterization of detector noise and GW signals, including IMBHs. We validate this approach by analyzing strain data from the third (O3) and fourth (O4) observing runs of the LIGO detectors. We used t-Distributed Stochastic Neighbor Embedding (t-SNE), an unsupervised clustering technique, to visualize the AST-derived embeddings of signals and glitches, revealing well-separated groups that align closely with independently validated Gravity Spy glitch classes. Our results indicate that the inductive bias from audio pre-training allows superior feature extraction compared to traditional supervised techniques, offering a robust, data-efficient pathway for discovering new, anomalous transients, and classifying complex noise artifacts in the era of next-generation detectors.

Figures (6)

• Figure 1: Overview of the proposed AST-LoRA glitch characterization workflow. (1) Input strain data is converted into log-mel spectrograms of shape ($1024 \times 128$). (2) The spectrograms are encoded by the AST encoder, fine-tuned via LoRA, to extract rich 768-dimensional semantic embeddings. (3) The high-dimensional embeddings undergo dimensionality reduction via PCA. (4) The reduced features are projected into a 3D latent space using t-SNE, revealing distinct clusters of glitch morphologies. The panels on the right display representative embedding vectors corresponding to the centroids of the identified clusters (colored red, orange, blue, and green).
• Figure 2: Log-mel spectrograms of a Blip, Extremely loud and Whistle glitch (top) and a Low frequency blip, Fast scattering and Low frequency burst glitch (bottom).
• Figure 3: Top: t-SNE map (left) and Silhouette scores (right) of glitch embeddings obtained from off-the-shelf AST encoder. The embeddings were generated by an AST encoder with no exposure to glitch data during training. Botton: Same as the top figures, but with embeddings obtained from AST encoder fine-tuned using PEFT (LoRA) on O3 glitch data.
• Figure 4: Top: t-SNE map (left) and Silhouette scores (right) of embeddings for standard BBH injections and 10 glitch classes. The embeddings were generated by an AST encoder fine-tuned exclusively on O3 glitches, with no exposure to GW data during training. Bottom: t-SNE map (left) and Silhouette scores (right) of high-mass BBH injections (total mass $\sim$ 100–1000 M$_{\odot}$) and glitch embeddings obtained using the same model.
• Figure 5: Top left: 3D t-SNE visualization of AST embeddings for 21 glitch types across all three observing runs, colored by unsupervised agglomerative (hierarchical) clustering applied in the 2D/3D t-SNE space following the procedure of Ferreira & González (2025) Tabata_tSNE. Top right: The same embeddings colored by Gravity Spy classifications. Bottom: Cross-match between agglomerative cluster labels and the dominant Gravity Spy class(es) present in each cluster. Our model identifies 18 clusters (labels 0–17), reflecting mergers of morphologically similar or overlapping Gravity Spy categories (e.g., related low-frequency families), and splits of some labeled classes into sub-populations when the embedding space resolves distinct modes.