Neural network approach to preferred event selection for low-latency gravitational-wave alerts
Phys. Rev. D 114, 063044 – Published 22 September, 2026
DOI: https://doi.org/10.1103/hsh6-64f4
Abstract
The LIGO-Virgo-KAGRA Collaboration uses multiple independent search pipelines to detect gravitational waves, often resulting in multiple triggers (so-called g-events) for a single astrophysical event. These triggers are grouped into superevents, raising a critical question for multimessenger astronomy: which g-event provides the most accurate sky localization in low latency for neutrino and electromagnetic follow-up? Currently, the g-event with the highest signal-to-noise ratio (SNR) is selected, under the assumption that it should provide the best estimators of the source’s parameters, including its location on the sky. Analysis of simulated signals reveals that the highest-SNR event often does not yield the smallest searched area—the metric we use to quantify localization accuracy. We present a neural-network-based selector trained on simulated signals comprising binary black holes, binary neutron stars, and neutron star–black hole mergers to identify the g-event with the minimum searched area. The network uses information such as detector SNRs, false-alarm rate (FAR), and chirp mass from all triggers associated with each astrophysical event and is designed to be pipeline agnostic. Across all events recovered below a FAR threshold of one per five months, the network reduces the mean searched area by relative to the SNR-based selector. The two selectors agree on of superevents; in the remaining cases, the network outperforms SNR by more than for of superevents (68% of these contain a neutron star), while SNR outperforms the network by a comparable margin in another (67% of these contain a neutron star). The net mean improvement is therefore modest. Unlike traditional selectors, the neural network preserves the underlying distribution of pipeline contributions, avoiding systematic biases toward specific pipelines. The network can be trained in approximately one minute on a few thousand events, making retraining straightforward as detector noise and pipeline configurations evolve between observing runs. Selection itself is instantaneous, making it suitable for low-latency applications. These results demonstrate that machine learning can enhance multimessenger astronomy capabilities while maintaining fairness across detection pipelines. We recommend exploring this approach for future observing runs.
