gw-nomad 0.1.0

NOMAD: NOn-stationary Modulation-Aware Denoiser for sgn

NOMAD

NOn-stationary Modulation-Aware Denoiser — an sgn-based reimplementation of NonSENS (Vajente et al., PRD 101, 042003) streaming noise subtraction for gravitational-wave strain data.

NOMAD measures excess noise in a target strain channel from auxiliary "noise witness" channels that couple into strain through transfer functions whose amplitude is modulated in time by slow "modulation witness" channels. The noise estimate is a sum of terms alpha_i(f) * s_i(t), where each s_i is a modulated signal (noise witness × products of modulation witnesses) and each alpha_i is an optimal filter learned offline. The stationary case (modulation_order = 0) is plain multi-channel Wiener filtering.

NOMAD's primary production role is the noise measurement: the SGN calibration pipeline's CAL-NOLINES → CAL-CLEAN stage is a plain subtractor that needs some other process to compute what to subtract (in O4 this was the NonSENS front-end installation). nomad-subtract --measure-only --publish-noise is that process, publishing the noise-estimate channel to arrakis on witnesses alone. It can also do the subtraction itself and publish or frame the cleaned channel directly.

Status

• v0.1: streaming application of trained models as an sgn pipeline, reading witness and target channels from arrakis (or frame files) and publishing the noise estimate and/or subtracted strain back to arrakis (or frames).
• v0.2: native model training (nomad-train, scipy L-BFGS over the analytic cost/gradient — no TensorFlow), plus conversion of original nonsens pickles (nomad-convert-model).

Installation

pip install .            # runtime
pip install .[convert]   # + dill, for converting original nonsens pickles

Usage

Train a model (or convert an existing NonSENS pickle):

nomad-train --config lho_isc.yaml --data-source frames \
  --frame-cache O4.cache --start 1400000000 --end 1400000600 \
  --output lho_isc.h5

nomad-convert-model lho_isc.pickle lho_isc.h5   # legacy pickles

The training config's target defaults to {ifo}:GDS-CALIB_STRAIN_NOLINES: the noise estimate is consumed downstream of the line subtraction, so training against NOLINES keeps already-removed lines out of the CSD estimates. Set target in the config (or pass --target) to train against any other channel, e.g. raw GDS-CALIB_STRAIN — as in the original nonsens, the target is simply whatever channel you name. A minimal config:

ifo: H1
noise_witnesses: [ASC-DHARD_P_OUT_DQ, ASC-DHARD_Y_OUT_DQ, 60.0]  # numbers = synthesized lines
modulation_witnesses: []
fs: 1024
t_fft: 10
fband: [8, 256]
n_sos: 10
glitches: {band: [15, 20], threshold: 3.0e-20}

Production: measure the excess noise and publish the estimate for the calibration pipeline's CLEAN subtractor (no target channel needed):

export ARRAKIS_SERVER=grpc://...
nomad-subtract --model lho_isc.h5 --data-source arrakis --queue-timeout 30 \
  --measure-only --publish-noise --publisher-id nomad-h1 \
  --noise-channel-name H1:CAL-NOMAD_NOISE_ESTIMATE

Standalone cleaning — arrakis in, subtracted strain out:

nomad-subtract --model lho_isc.h5 --data-source arrakis --queue-timeout 30 \
  --publish-arrakis --publisher-id nomad-h1 \
  --output-channel-name H1:GDS-CALIB_STRAIN_NOMAD

Published channels must be pre-registered with the arrakis server under the given --publisher-id, at the target sample rate (float64). Note the noise estimate carries the model's FIR high-pass group delay (length/2 seconds) as stream latency; timestamps are true GPS times, so a timestamp-aligned downstream subtractor handles it naturally.

Run offline — frames in, frames out:

nomad-subtract --model lho_isc.h5 --data-source frames \
  --frame-cache O3.cache --start 1242441180 --end 1242443180 \
  --frame-output-path '{instruments}-{description}-{gps_start_time}-{duration}.gwf'

Channels rarely share a frame type (strain lives in HOFT frames, witnesses in raw R frames); pass each extra cache with --auxiliary-cache (repeatable). Cache entries are grouped by frame type, one file per type is probed for its channels, and each channel is read from whichever frame type provides it — a single concatenated mixed cache works too. This applies to nomad-train as well.

Development

hatch test           # run the test suite
hatch run check      # mypy

Golden-comparison tests against the original nonsens implementation are skipped unless NOMAD_GOLDEN=1 is set and nonsens is importable.