Fast Deep Learning–based Imaging for Synthetic Aperture Array: Application to DART

Abstract DAocheng Radio Telescope (DART) is a radio synthetic aperture array consisting of 313 antennas, dedicated to monitoring solar activities and observing cosmic radio sources. To meet the real-time demands of interferometric brightness temperature imaging, we propose a deep learning–based fast-imaging method (DLSI). This method takes dynamically projected baselines and visibility functions as joint inputs, reconstructing brightness temperature images through a single forward pass. Extensive experiments demonstrate that the DLSI achieves imaging quality on DART solar observation data comparable to traditional methods such as gridding combined with CLEAN while reducing computational time by 3 orders of magnitude. Furthermore, under 15%–40% random antenna failure scenarios, the network maintains robust imaging performance, demonstrating exceptional adaptability to hardware malfunctions and strong generalization capability.