Optimizing searches for gravitational wave bursts using coherent WaveBurst 2G

TL;DR

This paper addresses detecting generic gravitational-wave transients without strong waveform priors by advancing the coherent WaveBurst-2G pipeline with a fully ML-based ranking statistic $\eta_0^{\prime}$ derived from the XGBoost score $W_{\text{XGB}}$. Utilizing 14.8 days of LVK O3 data, it evaluates two- and three-detector networks across three search configurations: HL-burst, HLV-burst, and HL-BBH, verifying background statistics with a Poisson model over a wide range of IFAR thresholds. The study demonstrates that combining HL and HLV searches via a logical OR yields a measurable boost in detection reach (up to about 8% in visible volume) and that a CBC-informed HL-BBH search outperforms agnostic HL-burst for stellar-mass CBCs, albeit with restricted parameter space. These results refine data-analysis strategies for current and future LVK runs by showcasing robust background validation, complementarities between detector networks, and the potential of model-informed post-processing within a largely agnostic framework. They also lay groundwork for extending all-sky burst surveys to targeted sources and for broader adoption of the PyCWB toolkit in the gravitational-wave community.

Abstract

The most general searches for gravitational wave transients (GWTs) rely on data analysis methods that do not assume prior knowledge of the signal waveform, direction, or arrival time on Earth. These searches provide data-driven signal reconstructions that are crucial both for testing available emission models and for discovering yet-to-be-uncovered sources. Here, we discuss progress in the detection performance of the coherent WaveBurst second-generation pipeline (cWB-2G), which is highly adaptable to both minimally modeled and model-informed searches for GWTs. Several search configurations for GWTs are examined using approximately 14.8 days of observation time from the third observing run by LIGO-Virgo-KAGRA (LVK). Recent enhancements include a ranking statistic fully based on multivariate classification with eXtreme Gradient Boosting, a thorough validation of the statistical significance accuracy of GWT candidates, and a measurement of the correlations of false alarms and simulated detections between different concurrent searches. For the first time, we provide a comprehensive comparison of cWB-2G performance on data from networks made of two and three detectors, and we demonstrate the advantage of combining concurrent searches for GWTs of generic morphology in a global observatory. This work offers essential insights for assessing our data analysis strategies in ongoing and future LVK searches for generic GWTs.

Figures (14)

• Figure 1: Workflow and data flow within cWB--2G.
• Figure 2: Implementation of time-slides in cWB-2G for two detectors: schematic view of three adjacent data segments. Top: actual data streams. Middle: time-lag treating each data segment as a circular buffer. Bottom: time shift by one data segment.
• Figure 3: Distribution of duration and bandwidth of background triggers reconstructed by cWB-2G on time--lags of HL whitened data. Color scale shows fraction of counts per bin, that are uniform in log scale. The red dashed line shows the Heisenberg--Gabor limit for the time-frequency representation of triggers.
• Figure 4: Distribution of reconstructed duration and bandwidth for the injected signals, zero--noise and whitened, used as training and testing sets in HL-burst and HLV-burst. The red dashed line shows the Heisenberg--Gabor limit. Color scale indicates fraction of counts per bin. Panel (a): training set of WNBs. Panel (b): testing set of ad--hoc waveform morphologies - sine-Gaussians near the Heisenberg--Gabor limit, shorter--duration Gaussian pulses, WNBs with larger time-frequency volumes.
• Figure 5: Distribution of CBC signals used for training and testing the HL--BBH search, in the duration-bandwidth plane. Panel (a) shows the simulated signals, zero--noise and whitened, with BNS and NS-BH signals appearing at longer durations, $\sim$ 1s. Panel (b) shows the whitened signals found by cWB 2G, which are then passed to the XGBoost classifier. Color scale shows the fraction of counts per bin. The red dashed line indicates the Heisenberg--Gabor limit.