A Neural Network Approach to Preferred Event Selection for Low-Latency Gravitational-Wave Alerts

TL;DR

A neural network-based selector trained on simulated signals to identify the g-event with the minimum searched area -- a metric quantifying localization accuracy, demonstrating that machine learning can enhance multimessenger astronomy capabilities while maintaining fairness across detection pipelines.

Abstract

The LIGO-Virgo-KAGRA collaboration uses multiple independent search pipelines to detect gravitational waves, often resulting in multiple triggers (g-events) for a single astrophysical source. These triggers are grouped into superevents, raising a critical question for multimessenger astronomy: which g-event provides the most accurate sky localization for 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 systematic deviations from this expectation. In particular, a false-alarm rate (FAR)-based selector performs slightly better than the SNR-based method, but introduces pipeline biases. We present a neural network-based selector trained on simulated signals to identify the g-event with the minimum searched area -- a metric quantifying localization accuracy. The network uses information (detector SNRs, FAR, and chirp mass) from all of the triggers associated with each astrophysical source and is designed to be pipeline-agnostic. Our results show that the neural network outperforms both traditional selectors, achieving a mean searched area ~2% smaller than the SNR-based selector. Unlike FAR-based selection, 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 and performs event selection instantaneously, 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 implementing this approach for future observing runs.

Figures (7)

• Figure 1: Left panel shows the training loss (trained on 90% of the events recovered in MDC11 at a FAR < 2/day) while the right panel shows the validation loss (evaluated on the remaining 10% of the events).
• Figure 2: Plots showing the mean searched areas of the skymaps chosen for significant (darker and thicker lines) and all (lighter and thinner lines) events by different selection algorithms. NN (sig) refers to the neural network (NN) trained on significant events, and NN (all) refers to the neural network trained on all events. Best refers to the event with lowest searched area. The spread ($1\sigma$) in the mean searched area is due to the 10 test sets. The solid (dashed) line denotes the mean searched areas for the skymaps of significant (all) events chosen by highest SNR (blue) and lowest FAR (orange).
• Figure 3: The left (right) column comprises of the cumulative distribution of searched areas for significant (all) g-event skymaps selected by lowest FAR, by highest SNR, by lowest searched area, and the neural network (NN) trained on significant (all) events. The median searched areas have been reported in parentheses. The bands have been obtained by evaluating different preferred event selectors on the 10 test datasets. The first panel comprises of all source types and the second panel comprises of sources with neutron stars.
• Figure 4: Skymaps associated with various g-events corresponding to S230309en. The highest SNR event, G927743, has an SNR of 12.17 and a searched area of 776 sq. deg. G927726, selected by a neural network, has an SNR of 11.43 and a smaller searched area of 197 sq. deg. The event with the best localization, G927745, is recovered at an SNR of 11.90 and has a searched area of 53 sq. deg.
• Figure 5: Fraction of selected g-events per pipeline for different selection choices. Those labeled best correspond to the g-event with the lowest searched area determined, i.e. the ground truth. The left column has significant events, while the right column has all events. The first panel comprises of all source types and the second panel comprises of sources with neutron stars. The dashed line shows the average fraction of g-events per superevent recovered by each pipeline. This is equivalent to picking g-events at random from a superevent.