DeepClean — learning to hear through the noise

The premise

Gravitational-wave observatories measure changes smaller than a proton. Their sensitivity depends not only on detecting faint astrophysical events, but on understanding the environmental and instrumental noise surrounding them.

DeepClean reframes noise subtraction as time-series regression: use auxiliary witness sensors to predict the noise coupled into detector strain, then remove that prediction while preserving the underlying astrophysical signal.

The constraint became the experiment

The original study relied on auxiliary detector channels that are not broadly public. Rather than pretend the missing data did not matter, our team made that constraint explicit. We constructed a controlled simulation using cleaned public strain data, injected noise with known behavior, and non-overlapping training and test intervals.

That gave us a narrower but defensible question: under controlled conditions, which behaviors of DeepClean could we reproduce—and where would the reproduction break down?

A loss with two views of the signal

Time-domain mean-squared error rewards accurate sample-level reconstruction. An amplitude-spectral-density objective rewards physically sensible behavior across frequency. We varied their relative weight to study the tradeoff rather than treating the published loss as a black box.

The frequency-domain contribution acted like spectral regularization. Pure or near-pure MSE converged quickly but generalized poorly; a balanced objective better preserved the residual spectrum while removing injected linear noise.

What held—and what did not

For linearly coupled jitter noise, the model reproduced the expected subtraction behavior on an unseen 1024-second segment. For simulated nonlinear 60 Hz power-mains sidebands, the central component was reduced but residual structure remained.

That limitation is part of the result. Nonlinear subtraction depends on realistic coupling dynamics and sufficiently rich witness channels. A model with the right theoretical capacity cannot recover information its inputs do not contain.

Reflection

The project sharpened a principle I return to often: model evaluation begins with the data-generating process. Reproduction is not just matching an architecture—it is identifying which assumptions made the original result possible, then designing an experiment that exposes the difference.

This was a collaborative research project. Architecture implementation, experimentation, evaluation, interpretation, and reporting were shared across the team.