AresGW Model 1: Deep Learning GW Pipeline

• AresGW Model 1 is a deep learning pipeline for gravitational wave detection using a one-dimensional ResNet architecture with enhanced signal pre-processing.
• It integrates dual high-pass and low-pass filtering, hierarchical trigger classification, and ensemble averaging to improve sensitivity and effectively reduce false alarms.
• Comparative analysis shows that AresGW Model 1 outperforms traditional pipelines by detecting more events and providing robust astrophysical probability estimates.

AresGW Model 1 is a deep learning pipeline for gravitational wave detection based on a one-dimensional ResNet architecture. It operates on data from interferometric detectors such as LIGO and Virgo, enabling rapid, robust identification of binary black hole merger events through a combination of signal pre-processing, hierarchical trigger classification, and advanced statistical ranking. Enhanced noise suppression and numerically stable confidence estimation provide significant gains over traditional matched-filter pipelines in sensitivity and false alarm reduction, particularly for moderate to low-SNR events across diverse observational data periods.

1. Architectural Enhancements and Signal Pre-processing

AresGW incorporates substantial modifications to its signal handling and neural pipeline. The whitening stage now utilizes both a high-pass filter (as in inverse spectrum truncation) and an added low-pass filter with a cutoff at 350 Hz. This dual filtering applies to both the training dataset and real-time analysis, greatly reducing the impact of high-frequency noise that previously compromised sensitivity to weak gravitational-wave signals.

Trigger detection is structured hierarchically. Candidates are assigned to three primary classes:

• Default Low-Pass: Identified with the standard low-pass filter at 350 Hz.
• Selective Noise Rejection: Triggers show significant change in the ranking statistic when the high-frequency cutoff is raised (400 Hz or 500 Hz), indicating proximity to noise artifacts.
• Selective Passband: Triggers stable under increased cutoff, representing the most robust candidates and delivering the lowest false alarm rate.

The raw output probability $\mathcal{R}$ from the network is transformed using double precision into a logarithmic ranking statistic,

$$\mathcal{R}_s = -\log_{10}(1 - \mathcal{R} + 10^{-16})$$

where ideal $\mathcal{R} = 1$ in theory yields $\mathcal{R}_s \to +\infty$, but practical FP64 limits bound $\mathcal{R}_s \leq 16$.

Ensemble averaging is performed by shifting the 1-second window by 0.001-second increments, averaging resulting $\mathcal{R}_s$ values ($\langle\mathcal{R}_s\rangle$) to reduce variability due to window placement and enhance overall trigger reliability.

2. Detection Algorithm and False Alarm Management

These architectural advancements lead to improved signal fidelity and enable AresGW to be more sensitive to lower-SNR events. By hierarchical trigger classification, frequency-dependent filtering, and numerical stability in $\mathcal{R}_s$, the model structure distinguishes genuine events from noise-induced triggers.

False alarm rates (FAR) are characterized by analytic fits, e.g., for the Default Low-Pass class:

$$\log_{10}(\text{FAR}) = -1.16\times10^{-4}(\mathcal{R}_s)^5 + 5.63\times10^{-3}(\mathcal{R}_s)^4 - 1.05\times10^{-1}(\mathcal{R}_s)^3 + 9.51\times10^{-1}(\mathcal{R}_s)^2 - 4.33\mathcal{R}_s + 9.69$$

Selective Passband triggers exhibit FAR reductions of one to two orders of magnitude relative to Default Low-Pass triggers. Elevating the high-frequency cutoff and comparing $|\mathcal{R}_s - \mathcal{R}_s'|$ enables effective noise rejection, while ensemble averaging maintains confidence stability across window shifts.

3. Comparative Performance and Sensitivity

AresGW outperforms conventional pipelines within its effective training region—component masses $7 \leq m_{1,2}/M_\odot \leq 50$, chirp mass $10 \leq \mathcal{M}/M_\odot \leq 40$. It confirmed 34 of 43 previously published GW events, while also identifying eight new candidate events, reaching a total detection count of 42. This surpasses the distinct event counts of established pipelines, such as mbta (27), pycbc_bb (31), gstlal (27), pycbc_broad (20), and a total distinct set of 36 among traditional approaches.

For Selective Passband triggers, the false alarm rate can drop to 0.0022/year at $\mathcal{R}_s=16$. Cumulative and inverse FAR plots demonstrate marked improvements, particularly at high $\mathcal{R}_s$ thresholds.

Pipeline	Distinct Events	Events in AresGW Band
mbta	27	Yes
pycbc_bb	31	Yes
gstlal	27	Yes
pycbc_broad	20	Yes
GWTC (distinct)	36	Yes
AresGW	42	Yes

4. Astrophysical Probability and Ranking

Astrophysical probability ($p_\text{astro}$) assessment relies on background and foreground modeling as functions of ensemble-averaged ranking statistic $\langle\mathcal{R}_s\rangle$.

Foreground cumulative rates are fit by power laws:

$F(x) = a(x - x_{\min})^b$

with background modeled similarly but incorporating inhomogeneous Poisson processes. The differential rates $f(\langle\mathcal{R}_s\rangle)$ (foreground) and $b(\langle\mathcal{R}_s\rangle)$ (background) determine

$$p_\text{astro} = \frac{f(\langle\mathcal{R}_s\rangle)}{f(\langle\mathcal{R}_s\rangle) + b(\langle\mathcal{R}_s\rangle)}$$

Signal injection studies into O3 data (as per the MLGWSC challenge) are used to calibrate ranking and FAR so that $\langle\mathcal{R}_s\rangle$ above set thresholds corresponds to $p_\text{astro} > 0.5$. The logarithmic ranking metric is robust against the extremely peaked values produced in high-confidence cases.

5. Candidate Event Analysis and Parameter Estimation

For new gravitational-wave candidate events, a comprehensive analysis workflow is deployed:

• Spectrograms: Time–frequency representations via Qp-transform methods visualizing the merger chirp signature.
• Parameter estimation: Using Bilby and the IMRPhenomXPHM waveform, parameters (chirp mass $\mathcal{M}$, $m_1$, $m_2$, mass ratio $q$, effective inspiral spin $\chi_\text{eff}$, luminosity distance $D_L$) are inferred, with posterior distributions illustrated in corner plots (Appendix A).
• Waveform reconstruction: Signal median waveforms (with 90% confidence intervals) are overlaid on bandpassed, whitened strain data; these reconstructions typically follow the data closely, even for low-SNR events, as shown in Appendix B.

New candidate events tend to exhibit lower SNR but are situated at greater luminosity distances than many cataloged events; their mass distributions remain consistent with previously characterized GW populations.

6. Empirical Validation and Robustness

Validation of AresGW spans synthetic background generation, empirical fits, and observational tests:

• A 10-year background yields analytic and empirical FAR agreement via time-shifts.
• Consistency checks enforce that inter-detector time delays do not violate causality, subject to uncertainty, and apply a $\chi^2$ waveform consistency test with frequency bin number $n \simeq 0.4[f_\text{peak}]^{2/3}$.
• Training and testing employ O3a noise from both Livingston and Hanford detectors; the model generalizes to Virgo-inclusive networks (LV, HV), even without explicit Virgo training.
• Testing on O1 and O2 periods recovers essentially all GW events, with ranking statistics near the theoretical maximum ($\langle\mathcal{R}_s\rangle \approx 16$).

Robustness is demonstrated via high-probability recovery of established GW events and detection of new candidates passing all physical and statistical consistency criteria.

7. Context and Prospects

AresGW’s integration of advanced pre-processing, shift-invariant trigger classification, double precision ranking, and ensemble averaging establishes a pipeline exceeding prior detectors in both sensitivity and computational efficiency. Its validated adaptability to multiple detectors and observing runs positions it as a strong candidate method for surge scenarios anticipated from next-generation interferometers. A plausible implication is the model’s suitability for real-time detection in lower-noise future runs, where rapid, fine-grained classification and robust parameter estimation will be essential for efficient gravitational wave astronomy.