🔭 astrophysics

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

This paper introduces a forward-modeling framework using the \texttt{pop-cosmos} generative model and machine-learned data models to infer KiDS-1000 redshift distributions directly from synthetic photometry, offering a robust alternative to spectroscopic reweighting that avoids selection biases and provides a critical cross-check for future Stage~IV cosmological surveys.

Original authors: Boris Leistedt, Hiranya V. Peiris, Anik Halder, Stephen Thorp, Daniel J. Mortlock, Arthur Loureiro, Justin Alsing, Gurjeet Jagwani, Madalina N. Tudorache, Sinan Deger, Joel Leja, Benedict Van den Buss

Published 2026-02-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Boris Leistedt, Hiranya V. Peiris, Anik Halder, Stephen Thorp, Daniel J. Mortlock, Arthur Loureiro, Justin Alsing, Gurjeet Jagwani, Madalina N. Tudorache, Sinan Deger, Joel Leja, Benedict Van den Bussche, Angus H. Wright, Shun-Sheng Li, Konrad Kuijken, Hendrik Hildebrandt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand the history of a massive city just by looking at a blurry, black-and-white photograph taken from a very high altitude. You can see the shapes of buildings and the general layout of streets, but you can't tell if a building is a new skyscraper or an old brick house, nor can you easily tell how far away it is.

This is the challenge astronomers face when studying the universe. They take pictures of millions of galaxies, but the most important piece of information—how far away they are (their redshift)—is often hidden or hard to guess just by looking at the image. Getting this distance wrong is like trying to calculate the size of a city but getting the scale wrong; your entire map of the universe becomes distorted.

This paper, titled "pop-cosmos: Forward modeling KiDS-1000 redshift distributions," introduces a new, smarter way to solve this puzzle. Here is the breakdown in simple terms:

The Problem: The "Guessing Game"

Traditionally, to figure out how far away galaxies are, astronomers have played a "matching game." They take a small sample of galaxies where they know the distance (because they used a powerful telescope to get a clear "spectrum" or fingerprint of the light) and try to match the blurry images of millions of other galaxies to this known sample.

The flaw: This known sample is often incomplete. It's like trying to guess the population of a whole country by only interviewing people who live in one specific, wealthy neighborhood. You might miss the poor, the elderly, or people in rural areas, leading to a biased map.

The Solution: Building a "Virtual Universe"

Instead of just matching blurry images to a limited list, the authors built a virtual universe from scratch. They call this framework pop-cosmos.

Think of it like this:

The Recipe (The Population Model): They created a digital "recipe book" for galaxies based on deep, high-quality data from the COSMOS survey. This recipe book doesn't just list ingredients; it understands the complex relationships between them. It knows that if a galaxy is a certain color, it likely has a specific age, mass, and dust content. It's like a master chef who knows exactly how changing the amount of salt affects the texture of a cake.
The Camera (The Data Model): A recipe is useless if you don't know how the camera takes the picture. The authors also built a "camera simulator" using machine learning. They trained it on millions of simulated images (called SKiLLS) that mimic the real KiDS telescope. This simulator learns exactly how the telescope blurs, dims, and distorts light, including the "static" or noise in the image.

How It Works: The "Forward" Approach

Most methods work "backward" (looking at a blurry image and guessing what it is). This paper works "forward":

Generate: They use the "Recipe Book" to create millions of perfect, theoretical galaxies with known distances and properties.
Simulate: They run these perfect galaxies through their "Camera Simulator." The simulator adds the blur, the noise, and the specific quirks of the KiDS telescope, turning the perfect galaxies into realistic, blurry images that look exactly like the real data.
Sort: They then take these simulated images and sort them into distance bins using the exact same computer programs the astronomers use for the real data.
Compare: Since they know the true distance of every galaxy they generated, they can now see exactly how the sorting program performs. They can say, "When the program says a galaxy is in the 'near' bin, it actually contains 5% of 'far' galaxies."

The Big Discovery: Two Different Recipes

To test their system, the authors tried two different "Recipe Books":

Recipe A (shark): A traditional, physics-based model that tries to simulate how galaxies form based on the laws of physics.
Recipe B (pop-cosmos): Their new, data-driven model that learned directly from real observations.

The Result: When they ran the simulation, the two recipes produced different maps of the universe. In the closest and farthest groups of galaxies, the average distance calculated by the two recipes differed by about 0.05 to 0.1.

To put that in perspective: In the world of cosmology, where we are trying to measure the expansion of the universe with extreme precision, a difference of 0.1 is like the difference between measuring a room as 10 feet wide versus 10.5 feet wide. It might seem small, but when you are calculating the energy of the entire universe, that small error adds up to a big mistake.

Why This Matters

The paper shows that the "Recipe Book" you choose matters immensely. The old physics-based recipe (shark) and the new data-driven recipe (pop-cosmos) tell slightly different stories about the universe.

The authors argue that pop-cosmos is likely more accurate because it learned directly from the real diversity of galaxies, rather than relying on simplified physics rules that might miss the messy details.

The Bottom Line:
This paper doesn't just give us a new map; it gives us a new tool to check our maps. By building a virtual universe and running it through a virtual camera, they can now predict exactly how accurate their distance measurements are, without relying on incomplete or biased real-world samples. This paves the way for future, ultra-precise surveys to measure the secrets of the universe without getting tripped up by the "blur" of our telescopes.

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

Problem Statement
The precision of cosmological constraints derived from Stage IV galaxy surveys is fundamentally limited by the accuracy of inferred galaxy redshift distributions, $n(z)$ . Traditional methods rely on re-weighting spectroscopic reference samples to match photometric data. However, these spectroscopic calibrations suffer from incompleteness and intrinsic selection biases, which can introduce systematic errors comparable to or exceeding the statistical uncertainties of modern weak lensing surveys. This paper addresses the challenge of inferring $n(z)$ for the Kilo-Degree Survey (KiDS-1000) weak lensing sample by reframing the problem as a forward-modeling exercise, bypassing the need for direct spectroscopic re-weighting.

Methodology
The authors develop a forward-modeling framework that synthesizes realistic mock observations of the KiDS-1000 survey. The framework integrates two primary components:

Generative Population Model (pop-cosmos): A score-based diffusion model trained on deep, multi-band photometry (26 bands) from the COSMOS2020 catalog (Spitzer IRAC Ch. 1 < 26). This model operates in the space of Stellar Population Synthesis (SPS) parameters (e.g., star formation history, metallicity, dust attenuation, AGN contribution) to generate realistic galaxy populations with intrinsic correlations between physical properties and observed photometry. Unlike semi-analytic models, pop-cosmos is data-driven and captures the full diversity of observed color–redshift relations.
Data Model (Machine-Learned): A set of machine-learning models trained on the SURFS-based KiDS-Legacy-Like Simulations (SKiLLS). These models learn the complex transformation from noiseless, intrinsic galaxy properties to realistic survey observations, including:
- Detection: A random forest classifier modeling the probability of detection based on $r$ -band magnitude and depth.
- Uncertainty: A conditional ODE flow modeling the uncertainty of the $r$ -band AUTO magnitude, which depends on morphology and local conditions.
- Noise: A conditional ODE flow modeling the 10-dimensional flux noise residuals (differences between measured and true fluxes) across all bands, capturing non-Gaussian tails and magnitude-dependent trends.
- Selection & Weighting: Additional classifiers and regressors emulate the KiDS-1000 quality cuts (e.g., star-galaxy separation, shape measurement flags) and predict lensfit weights.

The pipeline generates synthetic KiDS-like catalogs by drawing galaxies from the population model, applying the data model to simulate noise and selection, and then applying the same tomographic binning (using BPZ photometric redshifts) as the real data. The true redshifts of the simulated galaxies are then used as the calibrated $n(z)$ .

Key Contributions and Results

Validation of the Data Model: The authors demonstrate that their machine-learned data model can accurately reproduce the statistical properties of the SKiLLS simulations. When the SKiLLS underlying population (based on the shark semi-analytic model) is passed through the data model, the resulting redshift distributions match the "true" SKiLLS distributions with mean redshift differences $\Delta z < 0.005$ .
Impact of Population Assumptions: By replacing the shark population with the pop-cosmos population, the authors observe systematic shifts in the inferred redshift distributions. In the first ( $0.1 < z < 0.3$ ) and last ( $0.9 < z < 1.2$ ) tomographic bins, the mean redshifts differ by $\Delta z \sim 0.05\text{--}0.1$ .
Color–Redshift Relations: The shifts are attributed to differences in the color–redshift relations between the two models. pop-cosmos predicts a broader range of redder colors at high redshifts compared to shark, leading BPZ to assign intermediate-redshift galaxies to the highest bin and scatter some high-redshift objects to lower bins. This highlights that the choice of population model directly influences the calibration of $n(z)$ .
Generalization: The framework successfully applies a model trained on COSMOS2020 data to the KiDS-1000 survey, which has different depth, selection, and observing conditions. The resulting $n(z)$ estimates are compared to the publicly released KiDS-1000 distributions (derived via spectroscopic re-weighting), showing inherent differences due to the different underlying galaxy populations and selection functions.

Significance and Claims
The paper claims to provide a robust, independent alternative to spectroscopic-based redshift calibration. By forward-modeling the survey selection and noise directly from an empirically calibrated population model, the approach avoids the selection biases and incompleteness inherent in spectroscopic reference samples.

The authors emphasize that the observed systematic differences ( $\Delta z \sim 0.05\text{--}0.1$ ) between pop-cosmos and shark are comparable to the uncertainties reported for spectroscopic re-weighting methods. This underscores the critical importance of the underlying galaxy population model in determining $n(z)$ . The work establishes pop-cosmos as a viable tool for estimating population-level redshift distributions for Stage IV surveys. The authors state that these distributions will be used in an upcoming full KiDS cosmological reanalysis to provide an independent cross-check on existing calibrations. They conclude that as surveys push toward percent-level precision, forward-modeling tools grounded in empirically calibrated population models will be essential to ensure systematic uncertainties in redshift calibration do not limit fundamental physics constraints.

The Problem: The "Guessing Game"

The Solution: Building a "Virtual Universe"

How It Works: The "Forward" Approach

The Big Discovery: Two Different Recipes

Why This Matters

More like this