Deep learning-based imaging in radio interferometry

Abstract

Context. The sparse layouts of radio interferometers result in an incomplete sampling of the sky in Fourier space which leads to artifacts in the reconstructed images. Cleaning these systematic effects is essential for the scientific use of radiointerferometric images. Aims. Established reconstruction methods are often time-consuming, require expert knowledge, and suffer from a lack of reproducibility. We have developed a prototype deep learning-based method that generates reproducible images in an expedient fashion. Methods. To this end, we take advantage of the efficiency of convolutional neural networks to reconstruct image data from incomplete information in Fourier space. The neural network architecture is inspired by super-resolution models that utilize residual blocks. Using simulated data of radio galaxies that are composed of Gaussian components, we trained deep learning models whose reconstruction capability is quantified using various measures. Results. The reconstruction performance is evaluated on clean and noisy input data by comparing the resulting predictions with the true source images. We find that source angles and sizes are well reproduced, while the recovered fluxes show substantial scatter, albeit not worse than existing methods without fine-tuning. Finally, we propose more advanced approaches using deep learning that include uncertainty estimates and a concept to analyze larger images.