A deep learning approach to halo merger tree construction

Sandra Robles, Jonathan S. Gómez, Adín Ramírez Rivera, Nelson D. Padilla, Diego DujovnearXiv 2022 · DOI · OpenReview

TL;DR

This paper introduces a novel deep learning approach using Generative Adversarial Networks (GANs) to construct halo merger trees, which are crucial for understanding galaxy formation. Their GAN model can generate realistic merger trees with modest computational cost, outperforming traditional methods in certain aspects and offering a promising tool for large-scale cosmological simulations.

What Problem Does This Paper Address?

Understanding how galaxies form and evolve is a fundamental goal in astrophysics. Galaxies are believed to reside within dark matter (DM) halos, and their assembly history is intimately linked to the merger history of these halos. This history is typically encoded in "halo merger trees," which trace how smaller halos merge to form larger ones over cosmic time. The current gold standard for generating these trees involves computationally expensive, high-resolution N-body simulations. While these simulations are accurate, their cost limits the volume and number of simulations that can be run. Semi-analytic models (SAMs) of galaxy formation rely on these merger trees but are approximate in their physics. This paper addresses the need for a more computationally efficient yet accurate method to generate halo merger trees, bridging the gap between the precision of N-body simulations and the scalability of SAMs.

Methods

The core of this work is the application of Generative Adversarial Networks (GANs) to learn and generate halo merger trees. The authors adapted a previously developed GAN architecture, which uses convolutional neural networks (CNNs) for both the generator (encoder-decoder) and discriminator. The GAN is trained on existing merger trees from the EAGLE simulation suite.

Key methodological aspects include:

  • Input Variables: The GAN is trained using up to three input variables: the mass of halo progenitors and their final descendant, the progenitor type (main halo or satellite), and the distance of a progenitor to the main branch.
  • Tree Representation: Merger trees are represented in a matrix format, where rows correspond to snapshots (time steps) and columns to branches of the tree.
  • Training Strategy: The GAN is trained using a "minimax game" where the generator tries to create realistic trees that fool the discriminator, which is trained to distinguish real from fake trees. Reconstruction losses are added to improve the quality of generated trees.
  • Evaluation: The quality of the generated trees is assessed using the Kolmogorov-Smirnov (KS) test to compare statistical distributions of various properties (mass gain/loss, progenitor type duration, distance to main branch) between real and generated trees.

Datasets

  • Training and Evaluation Data: Merger trees are extracted from the dark matter-only (DMO) version of the EAGLE simulation suite. Specifically, the authors use the largest DMO simulation (EAGLE100) with a volume of 100 co-moving Mpc.
  • Halo Finder/Tree Builder Algorithms: Two distinct combinations of halo finder and tree builder algorithms were used to generate the training datasets:
  • SUBFIND (halo finder) with D-TREES (tree builder)
    • ROCKSTAR (halo finder) with CONSISTENTTREES (tree builder)
  • Benchmark: The generated merger trees are compared against "real" merger trees from the EAGLE simulation, serving as the benchmark.

Key Insights and Ideas

  • GANs as a Viable Alternative: The study demonstrates that GANs can effectively learn the complex structure and statistical properties of halo merger trees, offering a computationally cheaper alternative to traditional N-body simulations for generating these crucial data products.
  • Importance of Input Variables: Including progenitor type and distance to the main branch, in addition to progenitor mass, significantly improves the accuracy of the generated merger trees, particularly for the SUBFIND-based trees. This highlights that these seemingly secondary properties are crucial for accurately capturing the mass assembly history.
  • SUBFIND vs. ROCKSTAR: The GAN model shows better performance in reproducing SUBFIND-based merger trees compared to ROCKSTAR-based ones. This is attributed, in part, to a larger number of SUBFIND trees available in the training dataset.
  • Temporal Resolution Trade-off: Reducing the temporal resolution (number of snapshots) of the merger trees (e.g., from 201 to 101) improves the quality of the GAN-generated trees, suggesting that a balance between temporal detail and computational feasibility can be struck.
  • Scalability for Low/Intermediate Mass Halos: The GAN framework is particularly well-suited for generating merger histories of low- and intermediate-mass halos, which are the most abundant in cosmological simulations and are critical for understanding galaxy populations.

Key Results

  • The GAN model successfully generates well-constructed merger tree structures with high temporal resolution.
  • When trained with up to three variables (mass, progenitor type, and distance), the GAN reproduces the statistical features of the real merger trees with good agreement.
  • The inclusion of progenitor type and distance significantly improves the learned representation of halo mass growth, especially for SUBFIND-like trees.
  • Comparing equally sized samples, the GAN-generated SUBFIND-like trees show better agreement with EAGLE simulation trees than ROCKSTAR-like trees.
  • The KS test results indicate that the GAN can accurately reproduce mass gain/loss distributions, progenitor type duration, and distances to the main branch, with varying degrees of success depending on the input variables and the halo finder used.

Limitations and Future Work

  • Memory Constraints: The maximum number of branches the GAN can learn is limited by memory resources, necessitating careful selection of training data or potentially more advanced network architectures.
  • ROCKSTAR Tree Reproduction: While the GAN performs well, reproducing the statistical features of ROCKSTAR merger trees is more challenging, possibly due to differences in tree structure or fewer training examples.
  • Halo Finder Issues: The paper acknowledges that issues with halo identification (e.g., main halo-satellite switching in SUBFIND) can be transferred to the learned representation.
  • Future Work: The authors suggest that their GAN framework can be extended to include more input variables (e.g., position, peculiar velocity) to further enhance the realism of generated merger trees. They also propose that training on larger simulation volumes could help alleviate the lack of rare massive halo examples. The generated merger trees can be readily used with SAMs to simulate galaxy formation and evolution for upcoming large surveys.

Summary updated

Abstract

A key ingredient for semi-analytic models (SAMs) of galaxy formation is the mass assembly history of haloes, encoded in a tree structure. The most commonly used method to construct halo merger histories is based on the outcomes of high-resolution, computationally intensive N-body simulations. We show that machine learning (ML) techniques, in particular Generative Adversarial Networks (GANs), are a promising new tool to tackle this problem with a modest computational cost and retaining the best features of merger trees from simulations. We train our GAN model with a limited sample of merger trees from the Evolution and Assembly of GaLaxies and their Environments (EAGLE) simulation suite, constructed using two halo finders-tree builder algorithms: SUBFIND-D-TREES and ROCKSTAR-ConsistentTrees. Our GAN model successfully learns to generate well-constructed merger tree structures with high temporal resolution, and to reproduce the statistical features of the sample of merger trees used for training, when considering up to three variables in the training process. These inputs, whose representations are also learned by our GAN model, are mass of the halo progenitors and the final descendant, progenitor type (main halo or satellite) and distance of a progenitor to that in the main branch. The inclusion of the latter two inputs greatly improves the final learned representation of the halo mass growth history, especially for SUBFIND-like ML trees. When comparing equally sized samples of ML merger trees with those of the EAGLE simulation, we find better agreement for SUBFIND-like ML trees. Finally, our GAN-based framework can be utilised to construct merger histories of low- and intermediate-mass haloes, the most abundant in cosmological simulations.
BibTeX
@misc{robles2022deep,
  author = {Sandra Robles and Jonathan S. Gómez and Adín Ramírez Rivera and Nelson D. Padilla and Diego Dujovne},
  title = {{A deep learning approach to halo merger tree construction}},
  year = {2022},
  eprint = {2205.15988},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.15988},
}
Figure