Recovering Protoplanetary Disk Structure from Dirty Images via Score-Based Diffusion Models

In this study, we tackle the problem of reconstructing radio-interferometric images of protoplanetary disks (PPD) using Score Based Diffusion Models. We train on simulated ALMA-like images of PPDs and evaluate on held-out simulations with non-uniform sampling and Gaussian noise. The approach recovers key structures—rings and gaps across a wide range of sampling levels. Importantly, it still reconstructs trustworthy images even when the interferometer samples only about 30\% of spatial frequencies, a regime where the ``dirty'' image is strongly blurred and contaminated by sidelobes.