Very long base interferometry: a network of WaveletFlow neurons produces more detailed images of M87 and HL Tau

Applied to RadioAstron and ALMA archives, the method is scaled to the number of pixels and provides more statistical information than the conventional CLEAN algorithm.

Very Long Baseline Interferometry (VLBI) combines distant radio telescopes to achieve record angular resolutions. Bayesian sampling methods make it possible to obtain reliable images, accompanied by their uncertainty, but they remain computationally expensive, even since the transition from classical MCMC methods to normalizing flow neural networks: the cost rises with the number of details to be reconstructed.

Lev Kats, Alexey Goldin and Andrey Andrianov adopted the WaveletFlow architecture to increase the number of pixels in the reconstructed image distribution and get around this scale problem. They tested it on simulated data, then on archival observations: those of the RadioAstron space project on the galaxy M87 and those of ALMA on the protoplanetary disk of HL Tau.

Each image obtained is of a fidelity comparable to that of existing methods, but the approach provides more statistical information than CLEAN, the reference algorithm, which could count for the study of compact objects such as active galactic nuclei. The work is published in MNRAS.