A neural network for estimating compact binary coalescence parameters of gravitational-wave events in real time

TL;DR

This paper introduces a neural network that rapidly estimates gravitational-wave event parameters with high accuracy, providing dynamic bounds that improve the efficiency of real-time data analysis.

Contribution

A novel quantile regression neural network model for real-time gravitational-wave parameter estimation with dynamic bounds and reduced computational cost.

Findings

Model accuracy over 90% across datasets

Bounds reduce likelihood evaluations by 9%

Potential to shorten parameter estimation times

Abstract

Low-latency pipelines analyzing gravitational waves from compact binary coalescence events rely on matched filter techniques. Limitations in template banks and waveform modeling, as well as non-stationary detector noise cause errors in signal parameter recovery, especially for events with high chirp masses. We present a quantile regression neural network model that provides dynamic bounds on key parameters such as chirp mass, mass ratio, and total mass. We test the model on various synthetic datasets and real events from the LIGO-Virgo-KAGRA gravitational-wave transient GTWC-3 catalog. We find that the model accuracy is consistently over 90% across all the datasets. We explore the possibility of employing the neural network bounds as priors in online parameter estimation. We find that they reduce by 9% the number of likelihood evaluations. This approach may shorten parameter estimation run times without affecting sky localizations.