Poster Presentation

Advancing cluster weak lensing analysis with deep learning under realistic systematics

Author(s): Hyosun Park (Yonsei University), M. James Jee (Yonsei University, UC Davis), Yongsik Jo (UNIST), Minchang Jeong (UNIST), Siyeol Jeong (UNIST), Taehwan Kim (UNIST)

Presenter: Hyosun Park (Yonsei University)

Upcoming high-resolution astronomical imaging is increasingly hindered by systematics such as spatially varying PSFs and correlated noise. These effects directly affect the accuracy of cluster mass measurements from weak gravitational lensing. We develop a Transformer-based image restoration model trained on paired galaxy images to perform deconvolution and denoising. In initial tests with fixed PSF and Gaussian noise, the model successfully recovers photometric and morphological features, achieving near-JWST quality from HST-like inputs. We then extend the model to handle spatially varying PSFs and realistic noise correlations, applying it to Subaru images of the massive cluster MACS J0717.5+3745. The improved images provide more accurate galaxy shape measurements and a higher signal-to-noise ratio for weak lensing. Our approach aims not for a specific measurement but to enhance the quality of imaging data at the source. As more high-resolution observations become available, this method can serve as a general preprocessing tool for various cluster studies, including lensing, galaxy evolution, intracluster light, and comparisons to simulations.