Fast glitch modelling

An autoencoder fitting a blip glitch

Gravitational-wave data contains excess noise that affects the estimation of astrophysical source parameters. In some cases, the excess noise looks nearly identical to the noise seen previously, enabling us to categorize this noise into different classes (so-called glitch classes).

In my previous work (see antiglitch), I showed that it is possible to model short-duration glitches using a quasi-physical model. Using JAX, the glitches can be fitted to the quasi-physical model which then allows us to remove these glitches.

In this project, I show that short-duration glitches like blips and tomtes can be also modelled using autoencoders. With an autoencoder, I am able to fit a glitch model to the data orders of magnitude faster than JAX, while keeping the same precision.

For more, have a look at my GitHub repository: link.

Ronaldas Macas
Data Scientist

Applying my experience in gravitational-wave astronomy and machine learning to data science problems.