Type: Poster
Session: Posters (Wednesday & Thursday)
Author: Iago Lopes

Abstract: Cosmological measurements with Rubin LSST data rely on photometric redshift (photo-z) measurements that should attain unprecedented precision and accuracy. Consequently, massive development efforts are underway within and outside Rubin Science Collaborations. In the Dark Energy Science Collaboration (DESC), a wide array of projects seek to investigate and optimize photo-z analysis choices for different applications, including galaxy cluster redshift estimation, galaxy clustering and weak gravitational lensing correlations (i.e. 3x2pt analyses), and many more. In particular, evaluating the performance of photo-z algorithms in cosmoDC2 simulations is a key step in preparing for LSST data arrival. In this work, we investigate the performance of one of the leading machine learning (ML) photo-z algorithms, FlexZBoost. We train and test on cosmoDC2 data sets, seeking to identify and characterize issues with photo-z estimation, focusing on ensemble redshift distributions needed for tomographic 3x2pt analyses. We evaluate the algorithm's performance in a range of galaxy sample definition scenarios, including different magnitude cuts, and variations in training set representativeness and sky distribution. We also investigate different code configuration choices and their impact on the individual and ensemble metrics. The ensemble redshift distributions measured in this analysis will be used in cosmological inference pipelines to assess potential biases in cosmological parameters caused by imperfect ensemble redshift characterization.