Unsupervised Super-Resolution and Analysis of Real Lensing Images

Strong gravitational lensing is among the most powerful observational probes of dark matter substructure. High-resolution lensing images encode the morphological fingerprints of different dark matter models. Most available lensing images today and expected from upcoming surveys such as Euclid and LSST are ground-based and limited in resolution by atmospheric seeing. Obtaining matched high-resolution counterparts from space-based instruments like HST for every ground-based image is impractical at survey scale. This project addresses two main challenges in the proposed ML4Sci topic. First, it explores a super-resolution pipeline that does not require matched real low-resolution and high-resolution lensing observations for training. Instead, the model is trained on simulated high-resolution lensing images constructed with realistic source morphologies and survey-calibrated degradations, while remaining unsupervised with respect to real telescope HR targets. Second, it tests whether using real galaxy morphologies and limited ablations, can reduce the realism gap introduced by purely parametric Sérsic source models.