Robustness of Deep Learning Models to Precession in Gravitational-Wave Searches for Intermediate-Mass Black Hole Binaries

Abstract

Gravitational-wave searches for signals of intermediate-mass black hole binaries are hindered by detector glitches, as the increased masses from stellar-mass systems hinder current generation detectors from observing the inspiral phase of the binary evolution. This causes the waveforms to strongly resemble glitches, which are of similar duration within a similar frequency band. Additionally, precession of the orbital plane of a binary black hole may further warp signal waveforms. In this work three neural network-based classifiers for the task of distinguishing between signals and glitches are introduced, with each following different training regimes to study the impact of precession on the classifiers. Although all classifiers show highly accurate performance, the classifier found to perform best was trained following the principle of curriculum learning, where new examples are introduced only after the mastery of easier preceding examples. This classifier obtains an accuracy of approximately 95% on a synthetic test set consisting of signals and glitches injected into coloured noise from the O3 LIGO Hanford power spectral density. The model is compared to matched filtering, the state-of-the-art in modelled gravitational-wave searches, and analysed in search of particular sensitivities to black hole binary parameters. It was found that while the classifier is affected by the total mass of a system, the prediction of a misclassification is most strongly determined by visibility through the signal-to-noise ratio. The analysis of the three classifiers demonstrates that precession is handled differently depending on the training regime, meaning the architecture is not fully robust to precession and advancements can be made through the development of training routines.