Princeton University Seminar in Survey Science

Recent results from KiDS-Legacy cosmic shear analyses show that refined modelling of systematic effects can reconcile previously low estimates of the structure growth parameter with Planck CMB predictions. This underscores the importance of accurate treatment of observational systematics, such as photometric redshifts, and improved marginalisation over nuisance parameters for accelerated inference in forthcoming cosmological surveys such as Euclid and Rubin. In this talk, I will present a novel machine-learning framework that leverages Self-Organising Maps (SOMs) to project high-dimensional galaxy colour space into two dimensions, identifying spectroscopically under-sampled regions. These are augmented with synthetic galaxy catalogues to build representative training sets for photometric redshift calibration. Our method achieves sub-percent accuracy in mean redshift estimates for both LSST-Y1 and LSST-Y10, while robustly modelling systematic effects such as photometric noise, tomographic binning bias, and spectroscopic selection. I will also introduce the use of Variational Autoencoders (VAEs) for non-linear, lossless compression of redshift distribution realisations into a low-dimensional latent space. This enables efficient sampling and reconstruction of redshift covariances, supporting principled marginalisation over redshift distribution uncertainties. In addition, I will outline an analytic framework that reformulates angular power spectrum integration as tensor operations, accelerating computation on GPUs and enabling fully analytic marginalisation.