Speaker
Description
The unprecedented statistical power and imaging depth of the Vera C. Rubin Observatory's LSST will enable transformative tests of ΛCDM and the nature of dark matter and dark energy. Realising this potential hinges on accurately estimating ensemble redshift distributions and efficiently marginalising over their uncertainties.
I will present a new framework for photometric redshift calibration based on Self-Organising Maps (SOMs), which project the high-dimensional galaxy colour space onto a two-dimensional representation. This identifies regions poorly sampled by spectroscopy, which we then augment with synthetic galaxy catalogues to construct representative training sets. The method substantially reduces systematic biases for high-redshift galaxies, achieving sub-percent accuracy on the mean redshift for both Year 1 and Year 10 LSST configurations, while forward-modelling the dominant sources of systematic error — photometric noise, point-estimation bias, tomographic binning, and spectroscopic selection effects. Preliminary validation on early LSST DP1 data shows consistent performance, and a formal calibration analysis is in preparation for the upcoming DP2 release to support LSST anlyses on galaxy-halo connections.
Using these simulated n(z) ensembles, I show that conventional parametrisations — typically simple shifts along the redshift axis — can underestimate statistical uncertainties by up to an order of magnitude for galaxy clustering. To address this, I develop a data-driven approach based on a Variational Autoencoder (VAE) for non-linear dimensionality reduction, which outperforms standard PCA by compressing thousands of n(z) realisations into a Gaussianised latent space of substantially lower dimension. Resampling this latent space under Gaussian priors enables faithful reconstruction of the full n(z) covariance and accurate marginalisation within the cosmological likelihood.
Cosmological forecasts on synthetic LSST data show that weak-lensing constraints are robust across marginalisation choices, whereas the combination with large-scale structure reveals that conventional methods overestimate the dark energy Figure-of-Merit by 20% — underscoring the necessity of this improved framework for unbiased precision cosmology. The pipeline is being integrated into the LSST-DESC inference infrastructure for collaboration-wide application.