Untrained single scanning phase retrieval neural network for surface deformation measurement of large-aperture antennas

Abstract

Research on data-driven Fourier phase retrieval neural networks is well-established. However, in many practical applications, such as phase retrieval-based antenna surface deformation measurements in astronomy, obtaining a sufficient number of ground truth images for training is unlikely. In this paper, we propose a purely physics-driven neural network model for retrieving the aperture-field phase of large antennas. This is accomplished through the integration of the physical model describing the interaction between the aperture-field and the far-field with the neural network. Our simulations and experiments demonstrate accurate reconstruction of the aperture-field phase using only a single far-field intensity image. Furthermore, we present a solution tailored for scenarios with low signal-to-noise ratio.