Coherent WaveBurst (cWB) Pipeline

• Coherent WaveBurst (cWB) is a model-agnostic pipeline that uses time-frequency decomposition and maximum-likelihood inference to detect transient gravitational-wave signals.
• It employs Wilson–Daubechies–Meyer wavelets alongside machine learning and Gaussian mixture models to enhance sensitivity and suppress background noise.
• cWB has been pivotal in LIGO/Virgo campaigns, identifying key events like GW150914 and GW190521 while supporting multimessenger and targeted searches.

Coherent WaveBurst (cWB) is a model-agnostic, fully coherent analysis pipeline designed for the detection and reconstruction of transient gravitational-wave (GW) signals in interferometric data from global networks of GW observatories. Its architecture is based on a time–frequency analysis using Wilson–Daubechies–Meyer (WDM) wavelets and a constrained maximum-likelihood network statistic, enabling robust searches across a vast range of GW burst morphologies, including compact binary coalescences, core-collapse supernovae, and unanticipated transient sources. Since its initial deployment in 2003, cWB has been integral to every major observation campaign of the LIGO and Virgo detectors, serving as the primary burst discovery tool for signals such as GW150914 and GW190521. The pipeline implements advanced statistical inference, background estimation, and post-processing enhancements—most notably via machine learning and Gaussian mixture models—culminating in high-sensitivity, low-latency detection capability for both modeled and unmodeled GW events (Drago et al., 2020, Szczepanczyk et al., 2021, Szczepańczyk et al., 2022).

1. Algorithmic Principles and Time–Frequency Decomposition

The core of cWB is an excess-power, time–frequency (TF) analysis in the wavelet domain, applied to whitened strain data from each interferometer. The signal model assumes that the data vector in detector $k$ at TF bin $(i,j)$ is

$$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$

where $F_{+,k}, F_{\times,k}$ are antenna pattern coefficients for two GW polarizations and $n_k[i,j]$ is the Gaussian noise component (0802.3232). The WDM wavelet transform is used for its near-orthogonality and time–frequency localization. After normalization by the estimated noise, wavelet coefficients from all network detectors are coherently combined.

Pixels with excess normalized energy are selected, and neighborhood clustering yields candidate "triggers." Each cluster is subject to a constrained maximum-likelihood estimation, coherently reconstructing two waveform polarizations and maximizing the signal's likelihood across the whole detector array.

2. Coherent Statistic, Likelihood Formalism, and Vetoes

For a set of clustered pixels, cWB defines a fully coherent network likelihood statistic. In the Dominant Polarization Frame, the likelihood decomposes as

$$\mathcal{L} = \sum_{\Omega_{TF}}\left(\mathcal{L}_+ + \mathcal{L}_\times\right)$$

where closed-form maximum-likelihood estimators for $h_+, h_\times$ are obtained. The per-cluster (event) detection statistic is

$$\eta_c = \sqrt{\frac{E_c}{\max(\chi^2,1)}} \quad\text{with}\quad \chi^2 = \frac{E_n}{N_\text{df}},$$

where $E_c$ is the coherent energy and $E_n$ is the residual (null) energy, and $(i,j)$0 is the effective degrees of freedom (Szczepanczyk et al., 2021, Szczepanczyk et al., 2020).

Candidate triggers are subject to null-energy–based vetoes—e.g., requiring the network correlation coefficient $(i,j)$1 to exceed a threshold and reduced $(i,j)$2 to suppress incoherent transients. Only clusters passing these consistency conditions are considered as GW candidates (Mishra et al., 2024).

3. Search Configurations, Background Estimation, and Event Significance

cWB provides highly configurable search modes: all-sky, externally triggered (by gamma-ray, optical, or neutrino alerts), or specifically tuned to compact binary coalescences (CBC), including IMBH and BBH populations (Szczepanczyk et al., 2020, Mishra et al., 2024).

Statistical significance for each trigger is assessed empirically: time-shifted background trials (order $(i,j)$3 slides) are used to estimate the distribution of ranking statistics under the noise hypothesis, providing robust measurements of the false alarm rate (FAR) or its inverse (IFAR). Selection thresholds on $(i,j)$4 or derived statistics are set to achieve pre-defined FAR targets (e.g., one per year or per century) (Martini et al., 24 Oct 2025, Szczepańczyk et al., 2022).

4. Machine Learning and Post-Production Enhancements

Recent cWB developments supplant hard-threshold (binning) vetoes with supervised machine-learning classifiers:

• XGBoost: A gradient-boosted decision-tree ensemble, trained on cWB's multidimensional summary-statistic space spanning energy, correlation, waveform morphology, and blip-like features. The classifier outputs a penalty factor $(i,j)$5, which is multiplied by the coherent SNR to yield a final ranking statistic. At constant FAR, ML enhancement increases CBC detection efficiency by $(i,j)$6 and burst sensitivity by up to an order of magnitude in certain morphologies (Mishra et al., 2021, Szczepańczyk et al., 2022, Mishra et al., 2022).
• Gaussian Mixture Models (GMM): Multi-variate GMMs are trained on background triggers and synthetic signals, providing a log-likelihood ratio detection statistic $(i,j)$7 for each event. cWB-GMM is particularly effective for low-quality-factor (low-Q) burst signals and in suppressing blip glitches, showing up to $(i,j)$8 reduction in the amplitude threshold for detection at stringent FAR (Smith et al., 2024, Lopez et al., 2021, Smith et al., 1 Aug 2025).
• Autoencoders: Neural-network–based autoencoders, trained on specific glitch families (e.g., blip glitches), are folded into cWB as anomaly detectors. The reconstruction error is appended to the classifier feature vector, improving background suppression without loss of sensitivity to astrophysical signals (Bini et al., 2023).
• Specialized ranking for model-informed searches: cWB's architecture supports incorporation of model-informed features (e.g., chirp mass fits for BBH/IMBH) in the ranking/classification, enabling model-tuned searches with significantly improved efficiency for targeted source categories (Martini et al., 24 Oct 2025, Mishra et al., 2022).

5. Application Domains: CBC, Core-Collapse Supernovae, and Generic Bursts

cWB's model-agnostic approach has enabled detection and characterization of:

• Compact Binary Coalescences (CBCs): In the absence of matched-filter templates, cWB/ML achieves sensitivity and detection volume on par with, or exceeding, template-based pipelines for high-mass BBH and IMBH events. cWB recovered all known O1–O3 BBH mergers and identified BBH events missed by template searches, consistent with expectations from template mismatch studies (Mishra et al., 2024, Mishra et al., 2022).
• Core-Collapse Supernovae (CCSN): By making only minimal assumptions about burst morphology, cWB achieves a $(i,j)$9 detection efficiency distance up to $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$0 kpc for standard neutrino-driven models, reaching $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$1 kpc in rapidly rotating (MHD-driven) scenarios. Waveform reconstruction fidelity is excellent for short, narrowband signatures, with overlaps $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$2 at SNR $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$3 (Szczepanczyk et al., 2021, Gill et al., 2018). Recourse to Bayesian follow-up (e.g., BayesWave) and ML-based vetoes reduces background by $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$4–$$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$5 orders of magnitude at constant efficiency (Gill et al., 2018). Joint GW+neutrino detection scenarios further enhance identification probability (Sierra et al., 16 Apr 2026).
• Unmodeled and Unanticipated Bursts: cWB is natively robust against exotic morphologies, including cosmic-string cusps, memory bursts, magnetar flares, and higher-multipole harmonics. The burst search configuration, enhanced with GMM or ML ranking, offers the lowest effective amplitude threshold for single- or few-cycle signals—an essential capability for all-sky transient monitoring (Smith et al., 2024, Lopez et al., 2021).

6. Pipeline Architectures, Enhancements, and Open-Source Tools

The cWB codebase has undergone substantial architectural evolution:

• The second-generation pipeline (cWB-2G) introduced a fully ML-based ranking, expanded feature sets, improved time–frequency (multi-resolution) handling, and rigorous validation of statistical significance with Poisson-consistent background estimation. Detection volume significantly increased (by up to $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$6) over first-generation cWB at identical FAR (Martini et al., 24 Oct 2025).
• cWB XP incorporated a multi-resolution WaveScan/Gabor basis and a cross-power statistic (CRS) for improved time–frequency localization, substantially enhancing coherent background rejection, particularly in the context of multimessenger (GW+neutrino) CCSN studies (Sierra et al., 16 Apr 2026).
• Software frameworks: PycWB, a modular Python-based framework, exposes all cWB algorithmic stages—data conditioning, wavelet transforms, clustering, likelihood evaluation, ML/GMM ranking, and visualization—through compartmentalized modules with YAML-based configuration, error-checking, and multiprocessing or GPU acceleration (Xu et al., 2023).
• Denoising: rROF-based total variation denoising, implemented as a plugin, increases network SNR by up to $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$7 (iterative variant) in O1–O3 CBC events, translating to increased GW detectability and more precise source parameter inference without adversely affecting the FAR (Barneo et al., 2022).

7. Extensions: Higher Multipole, Population-Informed, and Multimessenger Searches

cWB's minimally-modeled framework can be reconfigured for targeted questions:

• Higher Multipole Searches: By selecting time–frequency strips corresponding to expected frequency–time tracks of $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$8 harmonics (e.g., $$w_k[i,j] = F_{+,k} h_+[i,j] + F_{\times,k} h_\times[i,j] + n_k[i,j]$$9), cWB can extract (3,3) modes in asymmetric mergers such as GW190814, providing model-agnostic confirmation of beyond-quadrupole GW emission (Halim et al., 2021, Vedovato et al., 2021).
• Population-Informed GMM: Training GMM classifiers on waveform populations of interest (e.g., dynamical black hole captures) yields targeted sensitivity, increasing range by up to $F_{+,k}, F_{\times,k}$0 and lowering astrophysical rate upper limits (Smith et al., 1 Aug 2025).
• Multimessenger (GW+EM/Neutrino): cWB operates in externally triggered and joint detection modes (e.g., using optical/EM or neutrino triggers), refining on-source windows, and integrating additional priors to improve identification probability and parameter estimation for events such as Galactic CCSN (Gill et al., 2018, Sierra et al., 16 Apr 2026).

---

• (Drago et al., 2020): Coherent WaveBurst, a pipeline for unmodeled gravitational-wave data analysis
• (0802.3232): Coherent method for detection of gravitational wave bursts
• (Szczepanczyk et al., 2021): Detecting and reconstructing gravitational waves from the next Galactic core-collapse supernova in the Advanced Detector Era
• (Gill et al., 2018): Enhancing the Sensitivity of Searches for Gravitational Waves from Core-Collapse Supernovae with a Bayesian classification of candidate events
• (Szczepańczyk et al., 2022): Search for gravitational-wave bursts in the third Advanced LIGO-Virgo run with coherent WaveBurst enhanced by Machine Learning
• (Mishra et al., 2022): Search for binary black hole mergers in the third observing run of Advanced LIGO-Virgo using coherent WaveBurst enhanced with machine learning
• (Mishra et al., 2021): Optimization of model independent gravitational wave search using machine learning
• (Lopez et al., 2021): Utilizing Gaussian mixture models in all-sky searches for short-duration gravitational wave bursts
• (Smith et al., 2024): Enhancing search pipelines for short gravitational wave transients with Gaussian mixture modelling
• (Bini et al., 2023): An autoencoder neural network integrated into gravitational-wave burst searches to improve the rejection of noise transients
• (Vedovato et al., 2021): Minimally-modeled search of higher multipole gravitational-wave radiation in compact binary coalescence
• (Halim et al., 2021): The search of higher multipole radiation in gravitational waves from compact binary coalescences by a minimally-modelled pipeline
• (Martini et al., 24 Oct 2025): Optimizing searches for gravitational wave bursts 2 using coherent WaveBurst 2G
• (Xu et al., 2023): PycWB: A User-friendly, Modular, and Python-based Framework for Gravitational Wave Unmodelled Search
• (Sierra et al., 16 Apr 2026): Joint Detection and Characterization of the Standing Accretion Shock Instability for Core-Collapse Supernovae with cWB XP
• (Barneo et al., 2022): Implementation of the rROF denoising method in the cWB pipeline for gravitational-wave data analysis
• (Smith et al., 1 Aug 2025): Search for dynamical black hole captures with Gaussian mixture modelling
• (Mishra et al., 2024): Gravitational Waves Detected by a Burst Search in LIGO/Virgo's Third Observing Run