Adam Zarka
Titre : Detection and characterization of galaxies with deep-learning in radio continuum surveys, preparation to SKAO
Équipe : GOSGAL
Encadrants : David Cornu, Benoit Semelin, Gregory Sainton
Site web : https://cianna.obspm.fr
The upcoming Square Kilometer Array Observatory (SKAO) will revolutionize the field of radio astronomy, particularly in the study of the Epoch of Reionization. However, its projected data flow (exceeding 700 PB/year) demands a shift toward exascale-ready pipelines. In addition, current automatic detection methods are beginning to be outpaced by the ever-increasing complexity of the data. In this Big Data era, Machine Learning stands out for efficient astronomical data processing.
The PhD aims to develop a robust and fast pipeline for radio-galaxy detection and characterization in 2D continuum images for SKAO, using precursor instruments (such as LOFAR, ASKAP, MeerKAT...) for methodological developments.
Building on the MINERVA team’s success with the YOLO-inspired (You Only Look Once) regression-based method during SKA Data Challenges 1 and 2, we generalize the method to train a new source detector model on a custom simulation pipeline with the aim of simulating any deconvolved radio interferometer response for full-field observations, including point sources, but also resolved Active Galactic Nuclei jets, star-forming galaxies, and ionospheric artifacts. The next step is to explore alternative self-supervised strategies like Denoising/Masked Autoencoders or Denoising Diffusion Probabilistic Models, pre-training generative networks on target data, in order to fully capture the specific features still missing in simulations.
