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. The prototype is delivered as a full-stack system: a PyTorch/FastAPI inference service, a Node.js orchestration API with caching, and an Angular command center for interactive residual visualization. On synthetic bursts injected into real LIGO noise, at a 1% false-alarm rate (about one false trigger per hundred noise-only tests), Keno residual excess power reaches near 100% detection efficiency from signal-to-noise ratio (SNR) 2 to 12. A deliberately wrong matched-filter template stays near ~20% efficiency on the same injections; a BBH-trained ResNet control evaluated off-template plateaus below saturation. On these synthetic bursts at 1% FAR, raw excess power without subtraction also saturates, so the injection campaign alone does not separate Keno from raw excess power. On published catalog GPS times the residual path can trigger where raw excess power does not (e.g. GW150914), and a coherent two-detector residual search recovers 9 of 18 dual-detector events after rejecting a known glitch near GW170817. Keno is meant to complement morphology-specific binary black hole classifiers, not replace them. The contribution is a reproducible software path for unmodeled residual search — model, API, and interactive tooling included. Source code: https://github.com/PanosM397/KenoFreeze label: 2026-07-paper-v1Checkpoint SHA256 prefix: 55ce7637…