Title:PIRATES -- a machine-learning framework for polarized, interferometric image reconstruction

Abstract:Optical interferometric image reconstruction is a challenging, ill-posed optimization problem which usually relies on heavy regularization for convergence. Conventional algorithms regularize in the pixel domain, without cognizance of spatial relationships or physical realism, with limited utility when this information is needed to reconstruct images. Here we present PIRATES (Polarimetric Image Reconstruction AI for Tracing Evolved Structures), the first image reconstruction algorithm for optical polarimetric interferometry. PIRATES has a dual structure optimized for parsimonious reconstruction of high fidelity polarized images and accurate reproduction of interferometric observables. The first stage, a convolutional neural network (CNN), learns a physically meaningful prior of self-consistent polarized scattering relationships from radiative transfer images. The second stage, an iterative fitting mechanism, uses the CNN as a prior for subsequent refinement of the images with respect to their polarized interferometric observables. Unlike the pixel-wise adjustments of traditional image reconstruction codes, PIRATES reconstructs images in a latent feature space, imparting a structurally derived implicit regularization. We demonstrate that PIRATES can reconstruct high fidelity polarized images of a broad range of complex circumstellar environments, in a physically meaningful and internally consistent manner, and that latent space regularization can effectively regularize reconstructed images in the presence of realistic noise.