Evenemanget har passerat

Exjobbspresentation, Victor Gustafsson

Machine Learning for Cosmic Magnetism: Preparing for the Square Kilometre Array Data Challenge


Evenemanget har passerat

Victor Gustafsson presenterar sitt examensarbete ”Machine Learning for Cosmic Magnetism: Preparing for the Square Kilometre Array Data Challenge”, utfört vid avdelningen för astronomi och plasmafysik, institutionen för rymd-, geo- och miljövetenskap. 

Handledare: Cathy Horellou

Opponenter: Olle Dahlberg, Gabriel Aspegren

Zoom link. (Password 230523)


The Square Kilometre Array (SKA) is an international megaproject to build the world'ss largest radio telescopes in Australia and in South Africa. The SKA and its precursors will generate unprecedented volumes of multi-dimensional data, necessitating advanced processing techniques. As such, the development of autonomous deep learning data analysis pipelines is essential to efficiently extract valuable insights from these vast datasets and further our understanding of the cosmos.

This study investigates the efficacy of deep learning (DL) techniques in detecting polarized signals in simulated data and recovering their input parameters. Polarization plays a crucial role in the field of astrophysics, particularly in the study of radio synchrotron radiation. It serves as a powerful tool for investigating magnetic fields within various celestial objects, including Active Galactic Nuclei (AGNs), which are galaxies housing supermassive black holes at their centers. Additionally, polarization allows us to examine magnetic fields along the lines of sight towards these objects. A technique called Faraday Rotation Measure Synthesis (RM synthesis) is employed to generate Faraday cubes, which involve applying a Fourier-type transform to data cubes containing polarization information across a broad frequency band. These cubes provide a multidimensional representation of the distribution of polarized intensities on the celestial sphere. In these Faraday cubes Gaussian and top-hat signals were simulated, subsequently transformed to frequency space, and subjected to Rician noise. Data preprocessing involved performing RM synthesis to create training data for a Convolutional Neural Network (CNN).

The samples underwent a two-stage process: initial classification as signal-containing or noise-only, followed by parameter estimation using a separate neural network. The classifier demonstrated near-perfect accuracy, with outliers predominantly observed in cases of low signal-to-noise ratios. Upon conducting a error analysis on the parameters of the signal, we found a Gaussian distribution centered around zero when disregarding the amplitude and width of the signals. However, when examining the predictions across the ranges of amplitudes and widths, a noticeable bias emerged.

In summary, our findings show that deep learning approaches hold significant potential for studies of polarization of astrophysical sources in radio astronomy.