[Paper Review] Demonstration of Machine Learning-assisted real-time noise regression in gravitational wave detectors

M. Saleem, Alec Gunny+18 more|arXiv (Cornell University)|Jun 20, 2023

Pulsars and Gravitational Waves ResearchPhysics and Astronomy3 citations

TL;DR

This paper demonstrates DeepClean, a convolutional neural network that enables real-time, low-latency noise regression in gravitational wave detectors by using witness sensors to estimate and subtract non-linear, non-stationary noise—such as 60 Hz power-line harmonics—improving signal-to-noise ratio without degrading astrophysical signals, with latency as low as 1–2 seconds.

ABSTRACT

Real-time noise regression algorithms are crucial for maximizing the science outcomes of the LIGO, Virgo, and KAGRA gravitational-wave detectors. This includes improvements in the detectability, source localization and pre-merger detectability of signals thereby enabling rapid multi-messenger follow-up. In this paper, we demonstrate the effectiveness of extit{DeepClean}, a convolutional neural network architecture that uses witness sensors to estimate and subtract non-linear and non-stationary noise from gravitational-wave strain data. Our study uses LIGO data from the third observing run with injected compact binary signals. As a demonstration, we use extit{DeepClean} to subtract the noise at 60 Hz due to the power mains and their sidebands arising from non-linear coupling with other instrumental noise sources. Our parameter estimation study on the injected signals shows that extit{DeepClean} does not do any harm to the underlying astrophysical signals in the data while it can enhances the signal-to-noise ratio of potential signals. We show that extit{DeepClean} can be used for low-latency noise regression to produce cleaned output data at latencies $\sim 1-2$\, s. We also discuss various considerations that may be made while training extit{DeepClean} for low latency applications.

Motivation & Objective

• To enable real-time, low-latency noise regression in gravitational wave detectors to improve detection sensitivity and enable rapid multi-messenger follow-up.
• To address the challenge of non-linear and non-stationary noise couplings—such as 60 Hz power-line harmonics—that degrade detector sensitivity.
• To validate that machine learning-based noise subtraction does not distort or harm genuine astrophysical signals in the data.
• To determine the optimal retraining frequency for the noise regression model to maintain performance over time.
• To demonstrate the feasibility of deploying DeepClean in production for upcoming O4 observing run with low-latency requirements.

Proposed method

• Utilizes a convolutional neural network (CNN) architecture called DeepClean to model and regress noise from gravitational wave strain data using auxiliary witness sensor channels.
• Trains the model on historical data from LIGO's third observing run (O3), using 1-second data frames to enable low-latency inference.
• Applies the trained model to clean real-time data by estimating and subtracting noise components, particularly 60 Hz and its sidebands, from the strain channel.
• Employs a strategy to mitigate edge artifacts in cleaned segments by waiting for the next 1-second frame, ensuring no edge effects in the current segment.
• Validates performance using downstream applications such as compact binary detection and parameter estimation on injected signals.
• Evaluates model stability over time by retraining on data from different days and measuring signal-to-noise ratio improvements via ASD ratios.

Figure 1: The top diagram illustrates the DeepClean architecture and the workflow. DeepClean takes timeseries data from multiple witness channels as input and runs it through a fully convolutional autoencoder. The autoencoder has four convolution layers for downsampling and four transpose-convolutio

Experimental results

Research questions

• RQ1 Can DeepClean achieve real-time noise regression with latencies of 1–2 seconds in gravitational wave detectors?
• RQ2 Does DeepClean improve the signal-to-noise ratio of astrophysical signals without introducing distortion or degradation?
• RQ3 How frequently must the DeepClean model be retrained to maintain optimal performance over time, especially for non-stationary couplings?
• RQ4 Can the model effectively suppress 60 Hz power-line noise and its sidebands arising from non-linear couplings?
• RQ5 What are the practical implications of deploying DeepClean in low-latency online processing for future observing runs like O4?

Key findings

• DeepClean successfully reduces 60 Hz power-line noise and its sidebands in real time with latencies of approximately 1–2 seconds.
• The signal-to-noise ratio of injected compact binary signals improved after DeepClean processing, with no degradation of the underlying astrophysical content.
• Model performance degrades when trained on data from earlier days and applied to later data, indicating that retraining every 1–2 days is necessary for optimal performance on 60 Hz noise.
• The ASD ratio of cleaned data exceeds 0.6 when using models trained on data from day 20, but remains below 0.6 when trained on earlier data, confirming time-varying coupling characteristics.
• The DeepClean model is computationally efficient and capable of retraining every 30 minutes or less, enabling frequent updates for non-stationary noise sources.
• The proposed edge mitigation strategy—waiting for the next frame—effectively replicates high-latency results in low-latency mode, validating the approach for online deployment.

Figure 2: This schematic shows the training strategy used for analyzing the mock data. The grey shaded segments represent science-quality data, and the yellow indicates that a model training is performed at the beginning of each science segment. The green segments represents one-hour long inference