Polarimetric, non-redundant aperture masking with next generation VAMPIRES: new instrumental capabilities, scientific outcomes, and image reconstruction techniques

Whilst many algorithms exist for interferometric image reconstruction, there are not yet algorithms for polarimetric interferometric image reconstruction. The polarisation state of light contains critical information otherwise uncaptured by standard, unpolarised interferometry, and many major facilities are now looking towards fully leveraging this information to broaden the observational reach of new and existing instruments. Polarimetric image reconstruction has additional challenges compared to unpolarised image reconstruction, as reconstructions of polarised images (Stokes I, Q and U) are spatial maps of vector components. As such, they need to individually and collectively display physically realistic and mutually consistent scattering physics. Within the present work, we demonstrate that a two-stage machine learning framework (a convolutional neural network (CNN) + iterative fitting) can be used to successfully perform polarimetric image reconstruction, whilst satisfying these challenging regularisation requirements. Using a custom set of MCFOST radiative transfer models, we train a convolutional neural network to learn the mapping between polarised images and interferometric polarimetric observables. We then deploy an iterative fitting mechanism inspired by the Deep Image Prior,¹ which iteratively improves the fit of polarimetric observables with cognisance of observational errors. In particular, the improvement provided by iterative fitting also results in the reconstruction of physically meaningful image structures that were missing from the original CNN image reconstruction. Our results suggest that this two-stage framework is a powerful tool for performing image reconstruction with complex regularisation constraints - in both polarimetric and non-polarimetric contexts. Here we briefly report our algorithm and initial results.