Characterising Dusty Star-Forming Galaxies Using Deep-Learning Super-Resolution Techniques

Single-dish observations of the far-infrared (FIR and sub-millimetre (sub-mm) regimes suffer from poor angular resolution and source confusion, complicating the study dusty star-forming galaxies (DSFGs). In this thesis, we build upon the super-resolution pipeline of Koopmans et al. (2025), adapting and extending its original scientific scope to deblend and denoise Herschel PACS 160 μm observations to characterise DSFGs. We employ a Swin-Unet transformer with a two-component loss function to recover spatial structure and source flux. The network is trained on a simulated data from SIDES/SHARK, using Spitzer IRAC 8 μm and MIPS 24 μm , and Herschel PACS 100 and 160 μm as input bands, injected with the associated instrumental noise and PSF FWHMs. The model’s performance is on simulations before application on COSMOS observations. The super-resolved COSMOS 160 μm fluxes are combined into a multi-wavelength catalogue, containing Euclid, COSMOS2020 and VLA 3 GHz, enabling SED fitting of the super-resolved DSFG subset. On simulations, the model is shown to improve the native resolutioon by a factor of 5, with astrometric accuracy of 0.′′25. Super-resolved fluxes underestimated by 7% for sources above ≥ 8 mJy . ≥ 95% reliability above ∼ 3 mJy and ≥ 90% cumulative completeness above 5 mJy indicate excellent performance above the total noise floor. When applied to the observed COSMOS 160 μm maps, the model effectively recovers source positions and fluxes, denoising and deblending the observed map. Super-resolved follow the general trend XID+ catalogue trend, but lie a factor of ∼ 4 below it, indicating a shortcoming of the source extraction process. Following SED fitting, determine the key physical properties of our DSFG subset, constraining the fraction of unobscured star formation as a function of stellar mass and redshift, and investigating the correlation between the total star formation rate and stellar mass – a relation dubbed the ’star-forming main sequence (SFMS).