Searches for Compact Binary Coalescence Events Using Neural Networks in LIGO/Virgo Third Observation Period

Abstract

We present the results on the search for the coalescence of compact binary mergers using convolutional neural networks and the LIGO/Virgo data for the O3 observation period. Two-dimensional images in time and frequency are used as input. The analysis is performed in three separate mass regions covering the range for the masses in the binary system from 0.2 to 100 solar masses, excluding very asymmetric mass configurations. We explore neural networks trained with input information from pairs of interferometers or all three interferometers together, concluding that the use of the maximum information available leads to an improved performance. A scan over the O3 data set, using the convolutional neural networks, is performed with different fake rate thresholds for claiming detection of at most one event per year or at most one event per week. The latter would correspond to a loose online selection still leading to affordable fake alarm rates. The efficiency of the neutral networks to detect the O3 catalog events is discussed. In the case of a fake rate threshold of at most one event per week, the scan leads to the detection of about 50% of the O3 catalog events. Once the search is limited to the catalog events within the mass range used for neural networks training, the detection efficiency increases up to 70%. A further improvement in the search efficiency, using the same kind of algorithms, will require the implementation of new criteria for the suppression of detector glitches.