Neural network method to search for long transient gravitational waves

We present a new method to search for long transient gravitational waves signals, like those expected from fast spinning newborn magnetars, in interferometric detector data. Standard search techniques are computationally unfeasible (matched filtering) or very demanding (suboptimal semicoherent methods). We explored a different approach by means of machine learning paradigms, to define a fast and inexpensive procedure. We used convolutional neural networks to develop a classifier that is able to discriminate between the presence or the absence of a signal. To complement the classification and enhance its effectiveness, we also developed a denoiser. We studied the performance of both networks with simulated colored noise, according to the design noise curve of LIGO interferometers. We show that the combination of the two models is crucial to increase the chance of detection. Indeed, as we decreased the signal initial amplitude (from ${10}^{\ensuremath{-}22}$ down to ${10}^{\ensuremath{-}23}$) the classification task became more difficult. In particular, we could not correctly tag signals with an initial amplitude of $2\ifmmode\times\else\texttimes\fi{}{10}^{\ensuremath{-}23}$ without using the denoiser. By studying the performance of the combined networks, we found a good compromise between the search false alarm rate (2%) and efficiency (90%) for a single interferometer. In addition, we demonstrated that our method is robust with respect to changes in the power law describing the time evolution of the signal frequency. Our results highlight the computationally low cost of this method to generate triggers for long transient signals. The study carried out in this work lays the foundations for further improvements, with the purpose of developing a pipeline able to perform systematic searches of long transient signals.