Published
Astrophysics

Keno

Generative noise subtraction for template-free gravitational-wave burst search

Keno is a research software prototype for searching short gravitational-wave bursts without assuming a known waveform shape. A neural network learns to predict detector noise from whitened sensor data; that prediction is subtracted; and the leftover residual is searched with an energy-based statistic, a two-detector timing check, and a simple consistency veto.

Keno

Deep dive

How it works

Generative noise subtraction

Instead of matched filtering against a known template, Keno learns what detector noise looks like and subtracts it — leaving a residual where unexpected signal energy can stand out.

  • Neural network predicts noise from whitened sensor data
  • Subtracted prediction yields a searchable residual stream
  • Energy-based statistic with two-detector timing check
  • Simple consistency veto to reject incoherent triggers

Full-stack research prototype

The project ships as a complete system — not just a model checkpoint — so results can be reproduced and explored interactively.

  • PyTorch model with FastAPI inference service
  • Node.js orchestration API with caching
  • Angular command center for interactive residual visualization
  • Documented freeze label and checkpoint for reproducibility

Evaluation highlights

Tested on synthetic bursts injected into real LIGO noise and on published catalog GPS times from GWOSC.

  • ~100% detection efficiency at 1% FAR for SNR 2–12 on synthetic injections
  • Wrong matched-filter template stays near ~20% on the same injections
  • Residual path can trigger where raw excess power does not (e.g. GW150914)
  • 9 of 18 dual-detector events recovered after glitch rejection near GW170817

Research positioning

Keno complements morphology-specific binary black hole classifiers — it is a reproducible path for unmodeled residual search, not a replacement for BBH pipelines.

  • Designed for short, unmodeled burst morphologies
  • Complements BBH-trained ResNet and matched-filter approaches
  • Contribution is the software path: model, API, and interactive tooling
  • CC BY 4.0 preprint and open-source repository

Interested in collaborating on research?

We welcome conversations about open science, reproducible software, and ambitious R&D prototypes.

Get in touch