Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AttenGW: A Lightweight Attention-Based Multi-Detector Gravitational-Wave Detection Pipeline

Victoria Tiki, Eliu Huerta

TL;DR

AttenGW addresses robust, low-latency multi-detector gravitational-wave detection by replacing graph-based detector fusion with a lightweight cross-attention module that fuses per-detector representations produced by a Hierarchical Dilated Convolutional Network. The end-to-end pipeline combines GWOSC-based preprocessing, a data generator with injected signals, and a PyTorch Lightning training/inference framework, totaling approximately $193{,}000$ trainable parameters. On real February 2020 O3b data and synthetic O3a data, a single AttenGW model achieves substantially lower background triggers than previous graph-based ensembles while maintaining competitive sensitivity, and a three-model ensemble matches or surpasses the six-model baseline. The work provides a reproducible, open-source software stack and demonstrates the practicality of attention-based aggregation for multi-detector GW searches, with future potential for longer signals, missing detectors, and online fine-tuning.

Abstract

We present AttenGW, an attention-based multi-detector gravitational-wave detection model and accompanying software stack designed for analysis of real LIGO data. AttenGW combines a per-detector hierarchical dilated convolutional network with an attention-based aggregation module that enforces cross-detector coherence, providing an alternative to graph-based aggregation schemes used in previous work. The pipeline adopts a LIGO-style preprocessing and data-loading workflow based on GWOSC time series, with standard whitening and filtering, and is released as a documented Python/PyTorch package. We benchmark AttenGW using simulated injections to estimate sensitive volume and on real O3 data, focusing on the February 2020 segment previously used to evaluate a spatiotemporal graph ensemble. On this month of data, a single AttenGW model reduces the false-positive rate relative to a single graph-based detector by a factor of a few, and an ensemble of three AttenGW models matches the performance of the corresponding six-model ensemble. Injection studies on real LIGO noise further indicate that attention-based aggregation yields stable performance on non-Gaussian backgrounds.

AttenGW: A Lightweight Attention-Based Multi-Detector Gravitational-Wave Detection Pipeline

TL;DR

AttenGW addresses robust, low-latency multi-detector gravitational-wave detection by replacing graph-based detector fusion with a lightweight cross-attention module that fuses per-detector representations produced by a Hierarchical Dilated Convolutional Network. The end-to-end pipeline combines GWOSC-based preprocessing, a data generator with injected signals, and a PyTorch Lightning training/inference framework, totaling approximately trainable parameters. On real February 2020 O3b data and synthetic O3a data, a single AttenGW model achieves substantially lower background triggers than previous graph-based ensembles while maintaining competitive sensitivity, and a three-model ensemble matches or surpasses the six-model baseline. The work provides a reproducible, open-source software stack and demonstrates the practicality of attention-based aggregation for multi-detector GW searches, with future potential for longer signals, missing detectors, and online fine-tuning.

Abstract

We present AttenGW, an attention-based multi-detector gravitational-wave detection model and accompanying software stack designed for analysis of real LIGO data. AttenGW combines a per-detector hierarchical dilated convolutional network with an attention-based aggregation module that enforces cross-detector coherence, providing an alternative to graph-based aggregation schemes used in previous work. The pipeline adopts a LIGO-style preprocessing and data-loading workflow based on GWOSC time series, with standard whitening and filtering, and is released as a documented Python/PyTorch package. We benchmark AttenGW using simulated injections to estimate sensitive volume and on real O3 data, focusing on the February 2020 segment previously used to evaluate a spatiotemporal graph ensemble. On this month of data, a single AttenGW model reduces the false-positive rate relative to a single graph-based detector by a factor of a few, and an ensemble of three AttenGW models matches the performance of the corresponding six-model ensemble. Injection studies on real LIGO noise further indicate that attention-based aggregation yields stable performance on non-Gaussian backgrounds.

Paper Structure

This paper contains 19 sections, 2 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Downloader
  4. Data generator
  5. Model
  6. Per-detector encoder (HDCN)
  7. Cross-detector aggregator (CAN)
  8. Output module
  9. Training
  10. Inference
  11. Results
  12. Sensitive volume curves on synthetic O3a data
  13. Training
  14. Inference
  15. Real O3b data and benchmark comparison
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Example training samples used for AttenGW. The plotted strain time series contain a binary neutron star signal generated with IMRPhenomPv2_NRTidal, injected into real LIGO O2 noise from August 2017 (Hanford and Livingston). The data are whitened and cropped to a fixed-length window; the shaded region marks the timesteps labeled as positive (merger neighborhood). Top row: scaled, high-SNR sample. Bottom row: Pure noise sample
  • Figure 2: HDCN block applied to the Hanford channel. Each layer applies the same processing steps, with the exception of the first one, where the initial convolution lifts the 1D input time series to 16 channels.
  • Figure 3: Aggregation modules: CAN block and Output module
  • Figure 4: