Leaving No Branches Behind: Predicting Baryonic Properties of Galaxies from Merger Trees

Chen-Yu Chuang, Christian Kragh Jespersen, Yen‐Ting Lin, Shirley Ho, Shy Genel

TL;DR

Researchers have developed a Graph Neural Network (GNN) that predicts complex galaxy properties—like stellar mass and color—by analyzing the entire "merger tree" history of dark matter halos. By leveraging the full assembly history rather than just a snapshot in time, the model outperforms standard machine learning methods by up to 45%, providing a high-speed "shortcut" for creating realistic mock universes.

What Problem Does This Paper Address?

In modern cosmology, we need massive "mock galaxy catalogs" to compare theoretical models with actual telescope data. However, we face a computational bottleneck:

  1. Hydrodynamical simulations (which model gas, stars, and black holes) are incredibly accurate but require millions of CPU hours.
  2. Dark Matter Only (DMO) simulations are much faster but don't actually contain "galaxies"—only the dark matter "halos" they live in.

The challenge is to find an accurate "mapping" that can paint realistic galaxies onto these dark matter halos. Previous attempts used simple Abundance Matching or basic Neural Networks that only looked at a halo's properties at a single point in time. These methods often missed the "memory" of the galaxy—how its history of mergers and accretion shaped its current state.

Methods

The authors present an upgraded version of a GNN model called Mangrove.

The Graph Approach

A galaxy’s history is a "merger tree": small halos merge to form larger ones over billions of years. The authors represent this as a directed graph where nodes are subhalos (with properties like mass and radius) and edges represent the progenitor-descendant relationships.

Key Innovation: Edge Rearrangement

Standard GNNs struggle with very deep graphs (like a 100-snapshot simulation) because information gets "washed out" (over-smoothing). To fix this, the authors introduced a Rearrangement of Edge Connections. They classified subhalos into categories (e.g., "progenitors," "post-merger," "pre-merger") and rewired the graph to prioritize information flow from significant merger events. This allows the network to "see" the impact of a major merger several billion years ago without needing a hundred layers of neurons.

Multi-Task Learning

The model doesn't just predict one value; it uses a combined loss function to:

  • Classify: Does this halo even contain stars or gas?
  • Regress: If it does, what is the specific value (e.g., stellar mass) and what is the uncertainty (variance)?

Datasets

The study utilizes the TNG300 simulation from the IllustrisTNG project.

  • Ground Truth: Data from the full Magnetohydrodynamical (MHD) simulation, which includes complex baryonic physics.
  • Input Features: Data from the corresponding Dark Matter Only (DMO) simulation.
  • Scope: The model was tested on galaxies with stellar masses $> 10^9 M_\odot$ across a wide redshift range ($z = 0$ to $5$).

Key Insights and Ideas

  • History Matters: The "branches" of the merger tree (the smaller halos that merged into the main one) contain vital information. Including them significantly improves predictions for stellar mass ($M_*$), gas mass ($M_{gas}$), and stellar metallicity ($Z_*$).
  • The Chaotic Limit: Some properties, like the Star Formation Rate (SFR) and galaxy color, are inherently "stochastic" or chaotic. The authors compared their results to the "chaotic uncertainty limit"—the point where even a perfect model can't predict better because the underlying physics is essentially random.
  • The Green Valley: The model is particularly good at identifying galaxies in the "green valley"—the transitional phase where galaxies stop forming stars and turn from blue to red—though predicting the exact timing of this transition remains a challenge due to its high stochasticity.

Key Results

  • Accuracy Boost: At redshift $z=0$, the GNN improved the scatter in stellar mass predictions by 45% compared to a standard Multi-Layer Perceptron (MLP) that ignores merger history.
  • High Correlation: The model achieved a Pearson correlation coefficient ($\rho$) of 0.979 for stellar mass, indicating an extremely tight fit to the "true" hydrodynamical data.
  • Robustness: The model maintains high accuracy across cosmic time ($z=0-5$), making it a versatile tool for studying the early universe.
  • Uncertainty Quantification: Because the model predicts a probability distribution (Gaussian likelihood), it "knows what it doesn't know," assigning higher uncertainty to galaxies with highly volatile star-formation histories.

Limitations and Future Work

  • Stochasticity Ceiling: For properties like SFR and color, the model is hitting a wall created by the "subgrid physics" of the simulation itself. No amount of merger tree data can perfectly predict these "jittery" properties.
  • Environmental Factors: The current model focuses on the internal history of the halo. Future versions could incorporate "environmental" data—like the density of the cosmic web surrounding the galaxy—to further refine predictions.
  • Sim-to-Real Gap: While the model emulates IllustrisTNG perfectly, it is still tied to that specific simulation's physics. The next step is applying these techniques to "constrained simulations" that aim to replicate our specific local patch of the real Universe.

Summary updated

Abstract

Abstract Galaxies play a key role in our endeavor to understand how structure formation proceeds in the Universe. For any precision study of cosmology or galaxy formation, there is a strong demand for huge sets of realistic mock galaxy catalogs, spanning cosmologically significant volumes. For such a daunting task, methods that can produce a direct mapping between dark matter halos from dark matter-only simulations and galaxies are strongly preferred, as producing mocks from full-fledged hydrodynamical simulations or semi-analytical models is too expensive. Here, we present a graph-neural-network-based model that is able to accurately predict key properties of galaxies such as stellar mass, g − r color, star formation rate, gas mass, stellar metallicity, and gas metallicity, purely from dark matter properties extracted from halos along the full assembly history of the galaxies. Tests based on the TNG300 simulation of the IllustrisTNG project show that our model can recover the baryonic properties of galaxies to high accuracy, over a wide redshift range ( z = 0–5), for all galaxies with stellar masses more massive than 10 9 M ⊙ and their progenitors, with strong improvements over the state-of-the-art methods. We further show that our method makes substantial strides toward providing an understanding of the implications of the IllustrisTNG galaxy formation model.
BibTeX
@article{chuang2024leaving,
  author = {Chen-Yu Chuang and Christian Kragh Jespersen and Yen‐Ting Lin and Shirley Ho and Shy Genel},
  title = {{Leaving No Branches Behind: Predicting Baryonic Properties of Galaxies from Merger Trees}},
  journal = {The Astrophysical Journal},
  year = {2024},
  doi = {10.3847/1538-4357/ad2b6c},
}
Figure