What Do a Malfunctioning Calibrator and a Thunderstorm Look Like in LIGO Data?

When LIGO Livingston recorded GW170817, a short burst of noise—a glitch—overlapped the signal. Scientists had to subtract it before they could infer the source's properties. The lesson has stuck: for gravitational-wave astronomy, noise is not a background detail. It is part of the measurement problem.

Glitches are short, non-Gaussian noise transients. In plain terms, they are short bursts that do not behave like the detector's ordinary background noise. They can come from the instrument or its environment, and new types can appear after detector changes. LIGO's detectors are laser interferometers built to detect gravitational waves. A loud enough glitch can hinder the identification or analysis of a real signal.

That is why the Gravity Spy project exists. Run on the Zooniverse community-science platform, it asks volunteers to classify spectrograms—visual maps of how power in the data spreads across time and frequency. Machine learning sorts large numbers of glitches into known classes, but it struggles when a new type appears. Most of the project's 37,000 volunteers classify known glitches. A smaller group hunts for new ones, using the Zooniverse Talk forum to compare examples and propose new classes.

A new preprint by researchers including members of the Gravity Spy collaboration examines two such proposals. It is a study of how community science can find rare features in LIGO data—and of why those features matter for detector characterisation, the work of understanding noise and its impact.

The glitch that looked like a meadow

The first proposal is called Photon Calibrator Meadow. It appears as a field of many flame-like transients below about 256 hertz. The proposal was made on 2 November 2024, using examples from LIGO Livingston data recorded on 18 August 2024.

Volunteers drew on public detector logs to form a hypothesis: the glitches were linked to a problem with the photon-calibrator system, a system used to calibrate the detector. The investigation supported that idea. The glitches were mostly confined to a period of less than one hour, matching a malfunction in the photon calibrator in the detector's Y-arm. Once the photon-calibrator issue was fixed, the glitches stopped.

Machine learning usually labelled these events as Fast Scattering (Crown) or Extremely Loud. Extremely Loud is an umbrella class for loud glitches with no single origin, so that label is not a problem. But contamination of the Fast Scattering class, which is linked to scattered light and ground motion, could complicate studies that rely on those classifications.

Because the photon-calibrator fault was fixed and the glitch was rare, Gravity Spy will not add this class yet. The team will watch for it to return.

A thunderstorm rumbling through the detector

The second proposal is called Vibration. It occupies mid-frequencies of roughly 30 to 200 hertz and lasts about one to four seconds. It has no simple shape; it often looks like a complicated, irregular network of peaks and horizontal lines. The proposal was made on 24 May 2019, using examples from LIGO Livingston during the third observing run. Similar examples appear in the fourth observing run and in LIGO Hanford data.

Volunteers identified three potential subclasses. Vibration Eruption includes higher-frequency bursts up to about 512 hertz, often at the start of the glitch. Vibration Burst lacks those higher-frequency peaks. Vibration Scattering shows horizontal structure similar to scattered-light glitches.

The subclasses often occur close together. For example, Vibration Eruption was often seen a few seconds before Vibration Scattering. That timing is similar to the roughly 12-second sound-travel time across one of the detector's arms. Thunderstorms are a known source of noise for the detectors. Depending on a thunderclap's distance, direction, and character—a sharp crack or a long rumble—acoustic and ground-motion disturbances can reach different parts of the detector at different times. Storms are more common in Louisiana than Washington, matching the greater number of examples at Livingston.

Cross-referencing a data-quality flag built from microphone data, the researchers found that 90% of the compiled Livingston examples coincided with thunderstorms. The remaining examples could be linked to human-made noise, such as liquid nitrogen deliveries, though that attribution is potential rather than settled.

Because this glitch is expected in multiple observing runs and is useful for detector characterisation, Gravity Spy will add Vibration as a single class.

When new glitches look like old classes

Both cases show how new glitch classes challenge automatic classification. The Vibration examples were mainly labelled as Scattered Light (Slow Scattering) or Scratchy, a confusion that reflects their similar shapes. The subclasses were classified differently: Vibration Eruption mainly as Scratchy, Scattered Light, and Extremely Loud; Vibration Scattering as Scattered Light; and Vibration Bursts as Fast Scattering, Scratchy, and Scattered Light.

The volunteers' collections can help retrain the classifier. That matters because glitches are not merely curiosities. They can hinder the search for gravitational waves and the analysis of signals already found. Monitoring data quality is an ongoing task, especially after detector changes.

The paper is a preprint and reports two volunteer proposals, not a complete inventory of LIGO's noise. The photon-calibrator case is not being added as a class for now because it was rare and tied to a fixed fault. The thunderstorm match applies to a compiled set of Livingston examples, not to every vibration-like event in every detector. And the causes of the remaining examples are not fully established.

Still, the result is a clear demonstration of what the Gravity Spy volunteers achieved. They noticed rare patterns in enormous data sets, connected them to detector and environmental records, and helped turn noise into knowledge. The next new glitch may already be hiding in a spectrogram, waiting for a human eye to see that it does not belong.