🔭 astrophysics

Learning the Universe with cosmological rescaling of merger trees and semi-analytic galaxy formation models

This paper introduces a computationally efficient method that applies cosmological rescaling directly to halo merger trees and semi-analytic models to generate galaxy populations across diverse cosmological and astrophysical parameters, achieving high-accuracy parameter estimation with negligible cost compared to traditional $N$ -body simulations.

Original authors: Richard Stiskalek, Lucia A. Perez, Shy Genel, Rachel S. Somerville, Raul E. Angulo, Sergio Contreras

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Richard Stiskalek, Lucia A. Perez, Shy Genel, Rachel S. Somerville, Raul E. Angulo, Sergio Contreras

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 entire history of the universe by watching a movie of how galaxies form. To get the physics right, scientists need to run massive, super-computer simulations. But here's the problem: running just one of these high-definition "movies" costs millions of dollars in computer time. If you want to test different theories about how the universe works (like changing the amount of invisible "dark matter" or the strength of cosmic expansion), you'd need to run thousands of these movies. That would cost more than the entire global economy could afford.

This paper introduces a clever shortcut, a "cosmic time-travel trick," that allows scientists to generate thousands of new universe simulations from just a handful of existing ones, saving massive amounts of money and time.

The Problem: The "Expensive Movie" Dilemma

Think of the universe as a giant, complex recipe. To see how the cake turns out, you have to bake it.

The Old Way: To test if adding a little more sugar (changing a cosmological parameter) changes the taste, you have to bake a whole new cake from scratch. If you want to test 1,000 different sugar levels, you need 1,000 separate baking sessions. This is what current "hydrodynamical" simulations do, and it's prohibitively expensive.
The Middle Way: Scientists developed "Semi-Analytic Models" (SAMs). Instead of baking the cake from scratch every time, they use a simpler set of rules to predict how the cake would look based on the ingredients. This is much cheaper, but it still requires a "base" cake (a simulation of dark matter) to start with. Even getting enough base cakes to test every possibility is too expensive.

The Solution: "Cosmic Rescaling"

The authors of this paper developed a way to take a single "base movie" of the universe and mathematically stretch or shrink it to look like a universe with different rules.

The Analogy: The Stretchy Rubber Sheet
Imagine the universe is drawn on a giant, stretchy rubber sheet.

The Original: You have a sheet with a specific pattern of dots (galaxies) and lines (gravity) representing our current universe.
The Trick: Instead of drawing a new sheet from scratch, you grab the corners of the existing sheet and stretch it.
- If you stretch it, the dots move apart, and the pattern changes.
- If you shrink it, the dots get closer together.
The Result: By stretching the sheet, you can make the original pattern look exactly like a universe where gravity is stronger, or where the universe is younger, without ever drawing a new line.

In the paper, they apply this "stretching" directly to merger trees.

What is a Merger Tree? Imagine a family tree for galaxies. It shows how small galaxies crashed into each other to form bigger ones over billions of years.
The Innovation: Previous methods tried to stretch the "pixels" (particles) of the universe, which was messy and inaccurate. This team figured out how to stretch the family tree itself. They take the history of how galaxies merged, mathematically adjust the timing and mass of those mergers, and then feed this "stretched" family tree into their galaxy-formation rules (the SAM).

The "Secret Sauce": Fixing the Stretch

When you stretch a rubber sheet, sometimes the patterns get distorted. The authors found that simply stretching the numbers didn't work perfectly; the "weight" of the galaxies (halos) came out slightly wrong, like a stretched photo that looks a bit blurry.

They invented a correction formula (based on a standard shape called an NFW profile) to fix this distortion. Think of it like a "smart filter" that automatically sharpens the image after you stretch it. With this single adjustment, they reduced the error in the galaxy weights to less than 1%, making the fake universe look almost identical to a real one.

The Results: One Cake Feeds a Thousand

The team tested this method using data from the CAMELS-SAM project (a large collection of universe simulations).

The Test: They took a small number of base simulations (as few as 64) and used their stretching trick to create 1,000 new, unique universe scenarios.
The Comparison: They compared these 1,000 "stretched" universes against 750 "real" universes that were simulated from scratch.
The Outcome: The "stretched" universes were just as good at predicting the number and distribution of galaxies as the expensive, from-scratch simulations.
- To get the same level of accuracy, the old method needed 750 expensive simulations.
- The new method needed only 64 expensive simulations, plus the cheap "stretching" math.

Why This Matters

This is like discovering a way to turn one loaf of bread into a thousand loaves without needing more flour or ovens.

Cost: Running the original simulation takes thousands of hours of computer time. Stretching the data takes a fraction of a second.
Efficiency: They showed that you can learn about the universe's secrets (like the amount of dark matter) just as accurately with this cheap method as with the expensive one.
Future: This allows scientists to explore "what-if" scenarios for the universe much faster, helping them understand the laws of physics without waiting years for computers to finish the work.

In short, the paper proves that you don't need to bake a new cake for every recipe variation. You can just stretch the dough you already have, fix the shape with a little magic, and get a perfect new cake every time.

Technical Summary: Learning the Universe with cosmological rescaling of merger trees and semi-analytic galaxy formation models

Problem Statement
Modern cosmology relies on precision measurements from large-scale structure surveys, necessitating accurate theoretical predictions across a wide range of cosmological and astrophysical parameters. While cosmological hydrodynamical simulations (e.g., Horizon-AGN, IllustrisTNG, FLAMINGO) are the gold standard for modeling galaxy formation, they are prohibitively expensive for the dense sampling of parameter space required by machine learning-based inference methods. Semi-analytic models (SAMs) offer a physically grounded, computationally cheaper alternative by evolving baryons along dark matter halo merger trees rather than solving hydrodynamics directly. However, even SAMs require underlying $N$ -body simulations to generate these merger trees. Running the hundreds or thousands of $N$ -body simulations needed for robust cosmological inference remains computationally prohibitive.

Methodology
This work extends the cosmological rescaling formalism (Angulo & White 2010) to operate directly on halo merger trees rather than particle snapshots or halo catalogues. The authors utilize the CAMELS-SAM suite, which couples $N$ -body simulations with the Santa Cruz Semi-Analytic Model (SC-SAM).

The core algorithm involves:

Tree-Level Rescaling: Instead of rescaling particles and re-running halo finders, the method rescales the properties of nodes within existing merger trees. The tree morphology (linkage between progenitors and descendants) is preserved, while halo properties are transformed to match a target cosmology ( $\Lambda'$ ) from an original cosmology ( $\Lambda$ ).
NFW-Based Correction: A novel correction is introduced to suppress systematic bias in rescaled halo masses, particularly when varying $\Omega_m$ and $\sigma_8$ . This correction modifies the rescaling of the NFW density profile parameters ( $\rho_0$ and $R_s$ ) using a free parameter $\alpha$ and a scale factor $A_{scale}$ derived from the virial overdensity threshold.
Time Interpolation: Since the CAMELS-SAM suite lacks snapshots at $z < 0$ , rescaling to lower $\sigma_8$ or $\Omega_m$ (where structure formation is slower) requires mapping the $z=0$ snapshot of the original simulation to a future redshift in the target cosmology. The method linearly interpolates halo properties between the two bracketing snapshots to obtain precise values at $z=0$ .
SAM Re-evaluation: The rescaled merger trees are processed through the SC-SAM to generate new galaxy populations. This allows for the generation of galaxy statistics across new cosmological and astrophysical parameters at negligible additional cost compared to running new $N$ -body simulations.

Key Contributions

Extension to Merger Trees: The authors demonstrate that cosmological rescaling can be applied directly to merger trees, bypassing the expensive halo-finding and tree-construction stages required by particle-level rescaling.
Bias Suppression: The introduction of the NFW-based correction (controlled by a single parameter $\alpha$ ) reduces systematic bias in rescaled halo masses to below the per-cent level for steps in $\Omega_m$ and $\sigma_8$ .
Validation: The method is validated against direct $N$ -body simulations with identical initial conditions but different cosmologies. The rescaling accurately reproduces individual halo properties (mass, concentration, spin) and galaxy properties (stellar mass, star formation rate) without significant systematic bias.
Parameter Estimation Efficiency: The authors apply the method to train neural network estimators for $\Omega_m$ and $\sigma_8$ using the Stellar Mass Function (SMF) and the two-point correlation function (2PCF).

Results

Accuracy: Rescaled merger trees produce galaxy populations that match the accuracy of direct simulations. The bias in rescaled halo masses is suppressed from $\sim 11\%$ (without correction) to $< 0.25\%$ . The scatter in stellar mass predictions from rescaled trees is comparable to the scatter introduced by varying the random seed in the SC-SAM satellite orbit model.
Computational Efficiency: Rescaling all merger trees from a single CAMELS-SAM $N$ -body simulation costs $\sim 0.1$ CPUh, compared to several thousand CPUh required to run the simulation itself.
Training Set Reduction:
- A training set of 1,000 rescaled simulations generated from as few as 64 base $N$ -body simulations matches the parameter estimation accuracy (RMSE) of a model trained on 750 dedicated $N$ -body simulations.
- Increasing the number of rescaled realisations to 3,200 (from 750 base simulations) improves the prediction of $\Omega_m$ by $\sim 25\%$ compared to training on the 750 base simulations alone.
- For the SMF, the performance is dominated by rescaling error, which is minimized with sufficient base simulations. For the 2PCF, performance is more sensitive to the number of independent initial condition phases (cosmic variance), saturating after $\sim 16$ unique phases.

Significance and Claims
The paper claims to demonstrate a practical route to obtaining predictions of galaxy summary statistics across cosmological and astrophysical parameters without the computational burden of running massive suites of $N$ -body simulations. By combining merger-tree rescaling with semi-analytic modeling, the authors show that it is possible to generate large, diverse training sets for machine learning inference with a fraction of the computational cost.

The authors emphasize that this approach acts as a "physics-informed interpolation scheme," allowing for the dense sampling of parameter space necessary for modern cosmological inference. They note that while the current results are modest regarding the specific observables (SMF and 2PCF) and the CAMELS-SAM suite, the method offers a scalable solution for future surveys (e.g., DESI, Euclid, Rubin) that require modeling of larger volumes and more complex parameter spaces. The work suggests that future suites could be optimized specifically for rescaling efficiency, potentially reducing the required number of base simulations even further.

The Problem: The "Expensive Movie" Dilemma

The Solution: "Cosmic Rescaling"

The "Secret Sauce": Fixing the Stretch

The Results: One Cake Feeds a Thousand

Why This Matters

More like this