Deep learning for clustering of continuous gravitational wave candidates

Abstract

In searching for continuous gravitational waves over very many ($\ensuremath{\approx}{10}^{17}$) templates, clustering is a powerful tool which increases the search sensitivity by identifying and bundling together candidates that are due to the same root cause. We implement a deep learning network that identifies clusters of signal candidates in the output of continuous gravitational wave searches and assess its performance. For loud signals, our network achieves a detection efficiency higher than 97% with a very low false alarm rate and maintains a reasonable detection efficiency for signals with lower amplitudes, i.e., at $\ensuremath{\lesssim}$ current upper limit values.