Accompanying Paper I, this sequel delineates a validation of the Artificial Intelligence for Regularisation in Radio-Interferometric Imaging (AIRI) algorithm on data from the Australian Square Kilometre Array Pathfinder (ASKAP). Here we showcase monochromatic sub-band AIRI-ASKAP images made from the same data via the same parallelised and automated imaging framework as used in Paper I: "uSARA validated on ASKAP data". Using a Plug-and-Play (PnP) approach, AIRI differs from uSARA by substituting a trained deep neural network (DNN) denoiser for the proximal operator in the regularisation step of the forward-backward algorithm during deconvolution. While the application of the DNN denoiser requires the computing power of a graphic processing unit (GPU), convergence in the AIRI deconvolution cycle is reached three times faster than in uSARA, thus significantly reducing computational cost. Using image domain noise estimates and peak flux values of dirty images as bounds of the target image dynamic range, we build a trained shelf of DNN denoisers that approximate prior models with comparable dynamic ranges of our selected data. Furthermore, we quantify variations of AIRI reconstructions when selecting the nearest DNN on the shelf versus using a universal DNN with the highest dynamic range. This Part II presents a validation of our AIRI-ASKAP images by comparing source structure, diffuse flux measurements, and spectral index maps to the results obtained in Paper I. Overall we see improvement over uSARA in the reconstruction of diffuse components. The most dramatic scientific potential with AIRI is perhaps inferred in the coverage, resolution, and accuracy of spectral index maps for our primary diffuse sources of interest: the radio phoenix in Abell 3395 and the X-shaped radio galaxy PKS 2014-558.
|Number of pages||10|
|Journal||Monthly Notices of the Royal Astronomical Society|
|Publication status||In preparation - 2022|