Denoising Interferometric Observations Using Visibility-Informed Neural Networks

Abstract The upcoming observations from the Square Kilometer Array Observatory will provide the astronomical community with a wealth of observations of important objects at long wavelengths. Full analysis of these outputs will necessitate specialized methods and software. Using synthetic observations of protoplanetary discs as an example, we present VIREO, a machine-learning-based visibility-informed method for denoising interferometric images. VIREO operates on image-plane observations, but it is informed by the interferometric measurement process through the UV-derived point spread function supplied as an additional input and used in the loss function. VIREO outperforms traditional cleaning methods and PSF-ignorant denoising models by producing data that is quantitatively cleaner and more conducive to analysis of the planets within the disc. Applying VIREO to archival ALMA data creates images with significantly less background noise, while maintaining, and in some cases enhancing, the substructure. By demonstrating the general utility of visibility-informed models, our results suggest that VIREO can be applied across interferometric observatories when trained on appropriate datasets.