pop-cosmos: Forward modeling KiDS-1000 redshift distributions using realistic galaxy populations

TL;DR

This work tackles the challenge of calibrating galaxy redshift distributions for Stage IV weak-lensing surveys by forward-modeling KiDS-1000 data through a combination of an empirically calibrated population model (pop-cosmos) and a data model learned from SKiLLS image simulations. By comparing pop-cosmos with the shark semi-analytic model under the same data model, it demonstrates that the choice of population can shift the tomographic redshift distributions by $\Delta z \sim 0.05$–$0.1$ in the edge bins, underscoring the importance of accurate color–redshift relations. The framework, trained on COSMOS2020 and SKiLLS, reproduces the observed data properties and provides a scalable, spectroscopic-calibration–independent path to $n(z)$ for KiDS-1000, with clear implications for Stage IV analyses. This forward-modeling approach offers a robust cross-check against spectroscopic calibrations and is extendable to full KiDS cosmology and future surveys where percent-level precision on cosmological parameters is demanded.

Abstract

The accuracy of the cosmological constraints from Stage~IV galaxy surveys will be limited by how well the galaxy redshift distributions can be inferred. We have addressed this challenging problem for the Kilo-Degree Survey (KiDS) cosmic shear sample by developing a forward-modeling framework with two main ingredients: (1) the \texttt{pop-cosmos} generative model for the evolving galaxy population, calibrated on \textit{Spitzer} IRAC $\textit{Ch.\,1}<26$ galaxies from COSMOS2020; and (2) a data model for noise and selection, machine-learned from the SURFS-based KiDS-Legacy-Like Simulations (SKiLLS). Applying KiDS tomographic binning to our synthetic photometric data, we infer redshift distributions in each of five bins directly from the population and data models, bypassing the need for spectroscopic reweighting. Keeping the data model fixed, we compare results using two different galaxy population models: \texttt{pop-cosmos}; and \texttt{shark}, the semi-analytic galaxy formation model used in SKiLLS. In the first ($0.1<z<0.3$) and last ($0.9<z<1.2$) tomographic bins we find systematic differences in the mean redshifts of $Δz\sim0.05$-$0.1$, comparable to the reported uncertainties from spectroscopic reweighting methods. This work paves the way for accurate redshift distribution calibration for Stage~IV surveys directly through forward modeling, thus providing an independent cross-check on spectroscopic-based calibrations which avoids their selection biases and incompleteness. We will use the \texttt{pop-cosmos} redshift distributions in an upcoming full KiDS cosmology reanalysis.

Figures (11)

• Figure 1: Validation of the $r$-band detection model: probability density of the GAaP magnitude limit given the noiseless model magnitude, for detected (left) and undetected (right) objects, for the SKiLLS simulations (black) and our model (red, described in described in \ref{['sec:detection-model']}). The contours enclose 68% and 95% of the total probability.
• Figure 2: Validation of the $r$-band uncertainty model: probability density of the AUTO flux uncertainty (vertical axes, converted to AB magnitude) as a function of the magnitude corresponding to a flux of $\sigma_{r_{\texttt{GAaP}}}$ (left) and the noiseless model magnitude (right), for the SKiLLS simulations (black) and our model (red, described in \ref{['sec:uncertainty-model']}). The contours enclose 68% and 95% of the total probability.
• Figure 3: Validation of the noise model: probability density of the flux signal-to-noise ratio as a function of the noiseless model magnitude in each band, for the SKiLLS simulations (black) and our model (red, described in \ref{['sec:noise-model']}). Contours enclose 68% and 95% of the joint probability density.
• Figure 4: Performance of the classifier predicting the KiDS-1000 class, as a function of the lensfit weight (true and predicted with the regressor presented in \ref{['sec:weight-prediction']}) and the photometric redshift. True positives show higher lensfit weights (left panel), confirming the classifier identifies high-quality sources. False positives cluster at low weights and high BPZ photometric redshifts $Z_B$, indicating contamination from low-quality measurements at the survey edges.
• Figure 5: Performance of the neural network regressor for lensfit weight prediction. Mean predicted weights (red) versus true weights (blue) as a function of AUTO magnitude, showing good agreement despite underestimated scatter.