In the era of high-precision cosmology, characterising the instrumental response of telescopes (PSF) has become a priority. Current and upcoming surveys such as Euclid, LSST or SKA will have unprecedented resolution, depth of field and coverage, enabling the detection of weak lensing signals and placing ever tighter constraints on cosmological parameters. To achieve this, these surveys will precisely measure the shapes of millions of galaxies, which are affected by the telescope's own PSF. To attain unbiased cosmological results, an accurate characterisation of the PSF is essential. During my PhD I apply ML and AI techniques to address key challenges in astronomical imaging, with particular emphasis on the instrumental characteristics of optical and radio telescopes.
WaveDiff is a wavefront-based, semi-parametric and data-driven PSF model that has been developed at CosmoStat in recent years. During my PhD, I contribute to the development and testing of new WaveDiff features, with the goal of applying the model to real Euclid observations. Furthermore, to enhance the capabilities of polychromatic PSF models such as WaveDiff, I proposed a method for stellar spectral classification from single wide-band images, increasing the number of stars available to fit the model. I also presented a new iterative optimisation strategy for the WaveDiff PSF model, which links its parametric and data-driven components, enabling the reconstruction of the PSF in wavefront space (WFE), making WaveDiff capable of reconstructing the WFE field solely from in-focus star observations.
PSF modelling challenges are not limited to optical imaging but also arise in radio interferometry, where signals from multiple antennas are combined to sample the spatial structure of the sky. In this context, the radio PSF is precisely determined by the antenna positions and the observing strategy, yet it exhibits a far more complex morphology than its optical counterpart. Imaging and deconvolution of interferometric observations have become a major challenge due to the enormous data volumes produced by modern radio interferometers. As part of my PhD, I collaborate with the ARGOS project, which aims to build a cutting-edge, affordable, and sustainable radio interferometer in Crete, Greece. To explore different antenna layouts and test reconstruction algorithms, I developed argosim, a Python package for radio interferometric simulations. The argosim package is lightweight, modular, and compatible with all major operating systems. It is built on JAX, providing significant performance acceleration, GPU compatibility, and automatic differentiation. We are currently investigating the use of argosim to explore antenna positioning optimisation, targeting specific PSF characteristics such as side lobe level, eccentricity, or size.
Next-generation interferometers, such as SKA, will observe the radio sky with unprecedented sensitivity and resolution. Their wide sky coverage will also enable weak lensing studies using radio data. Radio weak lensing not only complements optical observations but also allows access to higher redshifts. However, traditional shear estimation methods, based on measuring the ellipticity of observed galaxies, are not directly applicable to radio data. Interferometric observations are acquired in the Fourier domain, making shape measurements non-trivial. Furthermore, deconvolution methods such as CLEAN and its variants can introduce systematic errors, biasing shear estimates. To address these challenges, I propose a shear inference method based on forward generative modeling, which avoids explicit shape measurements. Our forward model simulates galaxy images, applies the shear transformation, and includes the instrumental response of the radio interferometer, with the likelihood explicitly defined in Fourier space. Using Monte Carlo sampling algorithm, the shear posterior can be estimated from galaxy observations.
