Machine learning for interferometric image reconstruction with sparse arrays

Interferometry, essential in radio and infrared astronomy, faces a significant challenge: reconstructing images from sparsely sampled data. Current regularized minimization algorithms rely heavily on predefined priors and hyperparameters, leading to ambiguities and inaccuracies in the images. Here, we present a project to integrate Neural Networks into interferometric image reconstruction. By utilizing the principles of Compressed Sensing and generative Neural Networks, this approach can map infrared interferometric data to reconstruct images more accurately, reducing reliance on rigid priors. The adaptability of the Neural Network ensures that the reconstructions are more precise and less dependent on user input, which is a significant advancement over current methods that require extensive expertise. In this work, we present, as software demonstration, reconstructions obtained from the Event Horizon Telescope data of the black-hole shadow at the core of M87.