Gravitational waves: a neural network combines all the detection pipelines from LIGO, Virgo and KAGRA and outperforms each of them individually

Rather than relying on a single algorithm, a classifier utilises the signals – or the absence thereof – from all the pipelines to better distinguish between noise and genuine events, and to reduce false alarms.

The search for transient signals of gravitational waves mobilizes several detection algorithms in parallel, sometimes redundant, sometimes complementary. A team including Erik Katsavounidis and Deep Chatterjee proposes to merge them using a neural network-based classifier, which takes into account what each pipeline has seen, or has not seen, when searching for coalescences of compact objects, in real time as well as in deferred analysis.

A more robust combined pipeline

Tested on simulated astrophysical signals injected into the real noise of the instruments, the method is evaluated using ROC curves. The resulting combined pipeline outperforms each pipeline taken in isolation as well as simple logic combinations. For the authors, the challenge is above all operational: more reliable detection in real time limits false alerts sent to telescopes around the world, simplifies the presentation of results and saves the resources of teams that follow events in multimessenger astronomy.