3D detection and characterization of ALMA sources through deep learning

ABSTRACT We present a deep learning (DL) pipeline developed for the detection and characterization of astronomical sources within simulated Atacama Large Millimeter/submillimeter Array (ALMA) data cubes. The is composed six DL models: convolutional autoencoder source spatial domain integrated cubes, Recurrent Neural Network (RNN) denoising peak frequency domain, four residual neural networks (ResNets) characterization. combination information improves completeness while decreasing spurious signal detection. To train test pipeline, we simulation algorithm able to generate realistic ALMA observations, i.e. both sky model dirty simulates always central surrounded by fainter ones scattered cube. Some were spatially superimposed in order deblending capabilities. performances compared those other methods significant improvements achieved. Source morphologies are detected with subpixel accuracies obtaining mean errors 10−3 pixel (0.1 mas) 10−1 mJy beam−1 on positions flux estimations, respectively. Projection angles densities also recovered 10 per cent true values 80 73 all set, While our fine-tuned data, technique applicable interferometric observatories, as SKA, LOFAR, VLBI, VLTI.