Advancing gravitational-wave searches from compact binary coalescences through novel algorithms and machine learning.

The first detection of Gravitational Wave event GW150914 by LIGO-Virgo-Kagra (LVK) collaboration opened the era of gravitational-wave astronomy. This discovery, occurring a century after Einstein’s prediction of gravitational waves in General Relativity, determined a scientific breakthrough. Since then, more than 170 events have been reported up to the first half of the current observing run (O4). These observations have enabled significant scientific advancements, including the first binary-neutron-star detection GW170817, the candidate intermediate-mass black-hole merger GW190521, and the most massive binary event to date, GW231123. Achieving such measures depends on continuous hardware upgrades and methodological improvements across the LVK detectors and data-analysis pipelines. Key challenges in detecting such extremely faint signals include noise sources characterization and isolation and developing algorithms that can distinguish transient noise from genuine astrophysical signals. In this context, we focus on data analysis for compact binary coalescences, using the Multi-Band Template Analysis (MBTA) pipeline, which employs matched-filtering to trigger signals in the interferometer data streams and it is currently deployed in observing runs. We investigate alternative strategies for MBTA template-bank construction that substantially reduce bank-generation time while maintaining high signal-recovery efficiency. We also assess a machine learning approach in which a Random Forests (RF) classifier is applied to MBTA triggers to separate noise from signal more effectively. Using O4a single-detector triggers, we compare RF-based statistics with respect to the MBTA ranking-statistics. We then estimate the astrophysical source probability pastro from the RF-derived statistics for coincident MBTA triggers in O3 and compare these values with the probabilities reported in the catalogues. Finally, we explore proof-of-concept Deep Learning methods for Gravitational Waves detection. We construct feature-rich representations from matched-filtering output statistics that can be learned by a Convolutional Neural Network, yielding an alternative triggering approach. This method offers advantages relative to the state-of-the-art pipelines: it requires a much smaller template bank and does not rely on the chi-sq consistency test to reject noise candidate triggers.