Simulating Transient Noise Bursts in LIGO with gengli

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "Simulating Transient Noise Bursts in LIGO with gengli".

Jocelyn: Here's the gist of the paper. Simulating transient noise burst with gengli develops a Generative Adversarial Network (GAN) to produce synthetic blip glitches in gravitational-wave interferometers, which is crucial for stress testing search pipelines and increasing detection confidence.

Vera: The most severe limitation to the detection of transient signals from astrophysical sources comes from transient noise artefacts, known as glitches, that happens at a rate around 1 per minute <ref:2205.09204#pg2>.

Jocelyn: Because glitches reduce the amount of scientific data available, there is a need for better modelling and inclusion of glitches in large-scale studies such as stress testing the search pipelines and increasing the confidence of detection <ref:2205.09204#pg2>.

Vera: The existence of gravitational-wave (GW) signals was successfully proven by LIGO and Virgo collaborations during the first observing run (O1).

Title and authors: Vera: So, this paper "Simulating Transient Noise Bursts in LIGO with gengli" is essentially presenting a method where they use a GAN to create realistic glitch data that mimics what we see in real gravitational wave observations >

Jocelyn: It's about using this system to generate known classes of glitches so they can better understand their features and morphologies, which is important for stress testing our pipelines >

Subrahmanyan: From a theoretical standpoint, being able to generate these synthetic signals helps us see how different types of noise interact with the detectors, which is crucial for structure formation studies <ref:2205.09204#pg1>

Vera: The core idea is that the generator learns the features of real glitch data so it can "fool" a discriminator network, which helps us create new samples that look like real glitches >

Jocelyn: They train this model by selecting blips from Livingston and Hanford that have a confidence score greater than or equal to zero point nine during the second observing run >

Subrahmanyan: It’s important to note they are using BayesWave to fit and reconstruct the input signal with datadriven wavelet models before feeding it into the network >

Vera: They then crop this reconstructed output to nine hundred thirty-eight data points, which is just a way of removing irrelevant data from the training set >

Jocelyn: And they evaluate that reconstruction quality by injecting it into real whitened noise and checking it with Gravity Spy to select glitches with a confidence score greater than or equal to zero point nine >

The paper's summary: Vera: The summary explains that the generator learns the features of real data so it can mimic those features, and they are using Wasserstein GANs because standard methods have problems with vanishing gradients and loss functions >

Jocelyn: They reformulate the optimization problem using the Earth’s mover distance or Wasserstein-one distance, which is represented as theta opt equals arg min theta w1(Px k Px˜) > <ref:2205.09204#pg1>

Subrahmanyan: That formulation is necessary because it handles the distribution between the true data distribution Px and the fake data distribution Px˜ in a way that's more stable for training <ref:2205.09204#pg3>

Vera: To make sure the discriminator doesn't get stuck, they add a gradient penalty term to the loss function, which results in l tot equals l φ θ plus lambda gp(φ) >

Jocelyn: Plus they use a consistency term to penalize points that are close to any observed real data point x, which helps keep the generated samples realistic >

Subrahmanyan: This whole setup is designed to allow the generator network to produce new samples by randomly taking a latent space vector and passing it through the generator >

Vera: So, in essence, they want this system to translate that random vector into real data that looks like a glitch without having to manually design every single noise pattern >

The paper's improvements: Jocelyn: The paper highlights that this generation process allows them to create raw glitches defined as the excess power of a whitened time series evaluated on a fixed length time grid at four thousand ninety-six Hertz >

Vera: They also emphasize that the amplitude of these generated glitches is arbitrary, so they have to scale them before injecting them into noise to achieve a target signal-to-noise ratio rho >

Subrahmanyan: And for filtering these anomalies, they consider three measures of distances: Wasserstein distance dW, mis-match dmm, and cross covariance dcc >

Jocelyn: The triplet of pW, pmm, and dcc is what they use as the novelty measure to filter glitches based on an anomaly score interval between pmin and pmax >

Vera: The open source package gengli is presented as a flexible tool that lets users scale or resample a glitch output easily >

Subrahmanyan: But they also point out a limitation, which is that the model's performance was assessed in a companion paper where several similarity distances indicated the model learned the underlying distribution of blip glitches despite anomalies due to imperfections of the input data set >

Conclusion: Jocelyn: So, to wrap up, this paper on "Simulating Transient Noise Bursts in LIGO with gengli" shows a way to generate artificial time-domain blip glitches using a GAN algorithm >

Vera: The performance assessment showed that the model learned the underlying distribution of blip glitches even though there were imperfections in the input data set >

Subrahmanyan: This means it gives us a new way to create mock data challenges for studies on future gravitational wave detectors and helps improve glitch classification with machine learning >

Jocelyn: The open source package gengli gives users an interface to scale and resample these generated glitches, which is really useful for testing the system >

Vera: We should keep an eye on this tool, especially as they plan to condition the GAN on other types of glitches like koi-fish and tomte using data from O3 >

Subrahmanyan: That will be very valuable because it enables the gravitational wave community to develop more realistic mock data challenges for studies on future GW detectors >