MADGRAV: a convolutional neural network chain detects 47 gravitational waves in LIGO data without a signal template

Rather than comparing the data with theoretical models, this pipeline detects anomalies. It is all the more effective when the colliding black holes are massive.

The reference method for detecting gravitational waves, the adapted filtering, confronts the data with a bank of theoretical signals. A team led by Gianluca Inguglia proposes a complementary route with MADGRAV, a sequence of convolutional neural networks that detect anomalies, classify instrumental artefacts, check the coherence between the two detectors of Hanford and Livingston and then prioritize the candidates, all on one-second windows transformed into time-frequency images.

Applied to data from the third observation campaign and the first two parts of the fourth, the pipeline retains 48 candidates above the threshold and claims 47 detections, with a false alarm rate of less than one per year and a probability of astrophysical origin of more than 90%. Forty-four of them are also found by cWB, the weakly modeled coherent search. The total masses of the systems range from 14 to 236 solar masses, for a median of 69.

A channel for the most massive mergers

The recovery rate rises significantly with mass: 8.1% of confirmed events in the catalogue under 30 solar masses, nearly 40% between 30 and 100, and 53.3% beyond. The authors see this as an independent detection channel, useful for cross-checking the appropriate filtering on massive mergers with a high signal-to-noise ratio.