A robust machine learning algorithm to search for continuous gravitational waves

Many continuous gravitational wave searches are affected by instrumental spectral lines that could be confused with a continuous astrophysical signal. Several techniques have been developed to limit the effect of these lines by penalizing signals that appear in only a single detector. We have developed a general method, using a convolutional neural network, to reduce the impact of instrumental artifacts on searches that use the SOAP algorithm Bayley et al. [Phys. Rev. D 100, 023006 (2019)]. The method can identify features in corresponding frequency bands of each detector and classify these bands as containing a signal, an instrumental line, or noise. We tested the method against four different datasets: Gaussian noise with time gaps, data from the final run of Initial LIGO (S6) with signals added, the reference S6 mock data challenge dataset Walsh et al. [Phys. Rev. D 94, 124010 (2016)] and signals injected into data from the second advanced LIGO observing run (O2). Using the S6 mock data challenge dataset and at a 1% false alarm probability we showed that at 95% efficiency a fully automated SOAP search has a sensitivity corresponding to a coherent signal-to-noise ratio of 110, equivalent to a sensitivity depth of $10\text{ }\text{ }{\mathrm{Hz}}^{\ensuremath{-}1/2}$, making this automated search competitive with other searches requiring significantly more computing resources and human intervention.