Fishing massive black hole binaries with THAMES

Hierarchical mergers in a dense environment are one of the primary formation channels of intermediate-mass black hole (IMBH) binary system. We expect that the resulting massive binary system will exhibit mass asymmetry. The emitted gravitational-wave (GW) carry significant contribution from higher-order modes and hence complex waveform morphology due to superposition of different modes. Further, IMBH binaries exhibit lower merger frequency and shorter signal duration in the LIGO detector which increases the risk of them being misclassified as short-duration noisy glitches. Deep learning algorithms can be trained to discriminate noisy glitches from short GW transients. We present the $\mathtt{THAMES}$ -- a deep-learning-based end-to-end signal detection algorithm for GW signals from quasi-circular nearly edge-on, mass asymmetric IMBH binaries in advanced GW detectors. Our study shows that it outperforms matched-filter based $\mathtt{PyCBC}$ searches for higher mass asymmetric, nearly edge-on IMBH binaries. The maximum gain in the sensitive volume-time product for mass ratio $q \in (5, 10)$ is by a factor of 5.24 (2.92) against $\mathtt{PyCBC-IMBH}$ ($\mathtt{PyCBC-HM}$) search at a false alarm rate of 1 in 100 years. Compared to the broad $\mathtt{PyCBC}$ search this factor is $\sim100$ for the $q \in (10,18)$. One of the reasons for this leap in volumetric sensitivity is its ability to discriminate between signals with complex waveform morphology and noisy transients, clearly demonstrating the potential of deep learning algorithms in probing into complex signal morphology in the field of gravitational wave astronomy. With the current training set, $\mathtt{THAMES}$ slightly underperforms with respect to $\mathtt{PyCBC}$-based searches targeting intermediate-mass black hole binaries with mass ratio $q \in (5, 10)$ and detector frame total mass $M_T(1+z) \in (100,200)~M_\odot$.