Skip to main content

Use of AI in Astronomy

Gravity Spy combines the crowd-sourcing power of citizen science with machine learning to classify glitches in gravitational-wave data. Gravitational-wave astronomy is a new, revolutionary method for observing the universe. Gravitational-wave detectors are extremely sensitive and complicated machines; to detect and analyse gravitational-wave signals requires a deep understanding of the detector noise. Glitches are transient bursts of noise, of a wide range of origins, which impact our ability to analyze gravitational-wave data. Classifying glitches allows us to hunt for their origins and mitigate their effects. Gravity Spy uses a convolutional neural net to classify glitches into classes based upon spectrogram images; citizen scientists perform a similar task, expanding our training set, and aiding classification of uncertain glitches. The machine learning algorithm allows us to rapidly classify glitches, producing near real-time reports for assessing candidate gravitational-wave detections, and enabling us to compile a large glitch database for study by detector experts. The Gravity Spy glitch classes are determined by detector experts, but as the detectors are commissioned, new glitches can arise. Gravity Spy empowers its citizen scientists to identify new glitch classes by allowing them to search for morphologically similar glitches to an interesting example. So far, volunteers have identified multiple new glitch types, demonstrating that members of the public, when provided with proper tools, can make meaningful contributions to cutting-edge science.

Spectrogram images showing the time–frequency structure of a selection of Gravity Spy glitch classes. The convolutional neural net classifies a set of four spectrograms showing different time intervals. Figure 1 of Bahaadini et al. (2018).

The Gravity Spy workflow is depicted above. Citizen scientists are trained up by progressing through a series of levels, each introducing new glitch classes and more uncertain classification. Examples where a consensus is reached are retired. New types of glitch can be searched for by looking for glitches close in feature space. Figure 2 of Coughlin et al. (2019).

Learn more about Northwestern's use of AI in research exploration

Back to top