Description
The exceptional statistical power and imaging depth of LSST will enable unprecedented constraints on the cosmological model, as well as the nature of dark matter and dark energy. However, accurately estimating ensemble redshift distributions and efficiently marginalising over their uncertainties remain major challenges. In this talk, I present a novel approach that leverages Self-Organising Maps (SOMs) to project the high-dimensional galaxy colour space into a two-dimensional representation. This allows us to identify regions of colour space poorly sampled by spectroscopy, which we then augment using synthetic galaxy catalogues to construct representative training datasets for photometric redshift estimation. Our method achieves sub-percent accuracy in the mean redshift distribution for both one-year and ten-year observations of Rubin and models realistic variations due to systematic effects. These results offer valuable insights for photometric redshift calibration using LSST early science data.
Using these simulated redshift distributions, we demonstrate that conventional parametrisations significantly underestimate statistical uncertainties—by up to an order of magnitude for galaxy clustering. To address this, we employ a Variational Autoencoder (VAE) to compress redshift distribution realisations into a Gaussianised latent space, enabling more efficient and accurate marginalisation. Cosmological forecasts show that weak lensing constraints remain robust regardless of marginalisation technique. However, when combined with large-scale structure datasets, our VAE-based method recovers a 20% downgrade in the Figure-of-Merit, correcting the overestimation from standard approaches. This highlights the necessity and effectiveness of our framework, which will be integrated into the LSST-DESC inference pipeline for collaboration-wide application in future precision cosmology analyses.
