A FIRST LOOK AT CREATING MOCK CATALOGS WITH MACHINE LEARNING TECHNIQUES

Xiaoying Xu, Shirley Ho, Hy Trac, Jeff Schneider, Barnabás Póczos, Michelle NtampakaThe Astrophysical Journal 2013 · DOI

TL;DR

This paper explores using machine learning (ML) techniques, specifically Support Vector Machines (SVM) and k-Nearest Neighbors (kNN), to predict the number of galaxies within dark matter halos. By avoiding pre-defined relationships like traditional HOD models, ML offers a more flexible and data-driven approach to constructing mock galaxy catalogs, crucial for analyzing large-scale cosmic structure.

What Problem Does This Paper Address?

Cosmologists rely on mock galaxy catalogs to simulate the universe and test their theories. These catalogs are essential for understanding the expected errors in observational data, calibrating analysis pipelines, and ultimately measuring fundamental cosmological parameters. Traditional methods for creating these mocks, such as Halo Occupation Distribution (HOD) models, often impose pre-defined, analytical relationships between halo properties and the number of galaxies they host. While effective, these assumptions might not fully capture the complex reality of galaxy formation. Furthermore, alternative methods like Subhalo Abundance Matching (SHAM) face challenges in accurately identifying subhalos within larger halos, a difficult task due to simulation resolution limitations. This paper aims to address these limitations by leveraging ML's ability to learn complex, non-parametric relationships directly from data, bypassing the need for explicit assumptions about the halo-galaxy connection and the complexities of subhalo identification.

Methods

The authors employ two machine learning algorithms for regression:

  • Support Vector Machines (SVM): Specifically, they use Support Vector Regression (SVR). SVMs work by finding an optimal hyperplane to separate data points in a high-dimensional space. For regression, SVR aims to find a function that best fits the data while allowing for a certain margin of error. They utilize a Gaussian (RBF) kernel, which allows the model to learn non-linear relationships.
  • k-Nearest Neighbors (kNN): This algorithm predicts the value for a new data point based on the average value of its 'k' nearest neighbors in the training data. The authors use a weighted average, where closer neighbors have a greater influence.

A key aspect of their approach is that these ML methods are non-parametric. This means they don't assume a specific functional form for the relationship between halo properties and the number of galaxies. Instead, the model is learned directly from the data. They also implement a feature selection algorithm (forward feature selection) to identify which halo properties are most predictive of the number of galaxies.

Datasets

The study utilizes data from the Millennium Simulation. This includes:

  • Halo Catalogs: Generated using a friends-of-friends (FoF) algorithm.
  • Semi-Analytic Galaxy Catalogs: Used to populate the halos with galaxies, providing information on galaxy properties like magnitudes and stellar masses.

The authors use a sample of 395,832 halos with masses greater than $10^{12}h^{-1}M_{\odot}$ for their primary analysis. They also create subsets for analyzing galaxy subpopulations based on color (blue/red) and stellar mass (high/low).

Key Insights and Ideas

  • Data-Driven Halo-Galaxy Mapping: ML offers a powerful alternative to traditional, assumption-laden methods for creating mock catalogs. By learning directly from simulation data, it can capture more complex relationships.
  • Circumventing Subhalo Identification: The ML approach, like HOD, relies only on parent halo properties, thus avoiding the difficult task of accurately identifying subhalos, which is a bottleneck for SHAM.
  • Feature Importance: The feature selection process reveals that the half-mass radius ($R_{1/2}$) is the most predictive halo property for the number of galaxies in the Millennium simulation, suggesting it encapsulates significant information about halo mass and environment.
  • Subpopulation Analysis: ML techniques can be readily extended to create mock catalogs for different galaxy subpopulations (e.g., by color or stellar mass), enabling studies of their distinct clustering properties.

Key Results

  • Predictive Accuracy: Both SVM and kNN achieved a Mean Squared Error (MSE) of approximately 0.16 for predicting the number of galaxies per halo ($N_{gal}$), which is significantly lower than the base MSE of ~0.505 (obtained by predicting the average $N_{gal}$ for all halos).
  • Distribution Matching: The ML predictions closely match the overall distribution of halos and the galaxy correlation function ($\xi(r)$) at large scales (within 5-10%).
  • Feature Selection: $R_{1/2}$ was identified as the most predictive halo parameter, yielding an MSE of 0.163, very close to the MSE obtained using all six features.
  • Subpopulation Performance: While ML shows promise for creating mocks of galaxy subpopulations, there's a slight underprediction of power at small scales, particularly for red galaxies and high stellar mass galaxies, which can impact studies of galaxy evolution. However, agreement at large scales remains good.

Limitations and Future Work

  • Underprediction Bias: The ML models exhibit a slight bias towards underpredicting $N_{gal}$, especially for halos with a large number of galaxies. This is attributed to the rarity of such halos in the training data and can lead to a deficit in clustering power at small scales.
  • Feature Completeness: The current features (halo mass tracers and environment indicators) may not fully capture the halo-galaxy mapping, suggesting that other factors like the full merger history of the halo could be important.
  • Future Directions: The authors suggest several avenues for future research:
  • Using larger training datasets to mitigate the underprediction bias.
  • Developing methods to penalize bias more heavily or give more weight to high $N_{gal}$ neighbors in kNN.
  • Extending ML to predict not only the number of galaxies but also their positions and velocities within halos.
  • Applying these techniques to more sophisticated cosmological simulations (e.g., hydrodynamical simulations) once they become available.

Summary updated

Abstract

We investigate machine learning (ML) techniques for predicting the number of galaxies (<em>N</em><sub> gal</sub>) that occupy a halo, given the halo's properties. These types of mappings are crucial for constructing the mock galaxy catalogs necessary for analyses of large-scale structure. The ML techniques proposed here distinguish themselves from traditional halo occupation distribution (HOD) modeling as they do not assume a prescribed relationship between halo properties and <em>N</em> <sub>gal</sub>. In addition, our ML approaches are only dependent on parent halo properties (like HOD methods), which are advantageous over subhalo-based approaches as identifying subhalos correctly is difficult. We test two algorithms: support vector machines (SVM) and <em>k</em>-nearest-neighbor (kNN) regression. We take galaxies and halos from the Millennium simulation and predict <em>N</em> gal by training our algorithms on the following six halo properties: number of particles, <em>M</em> <sub>200</sub>, σ <sub><em>v</em> </sub>, <em>v</em> <sub>max</sub>, half-mass radius, and spin. For Millennium, our predicted <em>N</em> gal values have a mean-squared error (MSE) of ~0.16 for both SVM and kNN. Our predictions match the overall distribution of halos reasonably well and the galaxy correlation function at large scales to ~5%-10%. In addition, we demonstrate a feature selection algorithm to isolate the halo parameters that are most predictive, a useful technique for understanding the mapping between halo properties and <em>N</em> <sub>gal</sub>. Lastly, we investigate these ML-based approaches in making mock catalogs for different galaxy subpopulations (e.g., blue, red, high <em>M</em> <sub>star</sub>, low <em>M</em> <sub>star</sub>). Given its non-parametric nature as well as its powerful predictive and feature selection capabilities, ML offers an interesting alternative for creating mock catalogs.
BibTeX
@article{xu2013first,
  author = {Xiaoying Xu and Shirley Ho and Hy Trac and Jeff Schneider and Barnabás Póczos and Michelle Ntampaka},
  title = {{A FIRST LOOK AT CREATING MOCK CATALOGS WITH MACHINE LEARNING TECHNIQUES}},
  journal = {The Astrophysical Journal},
  year = {2013},
  doi = {10.1088/0004-637x/772/2/147},
}
Figure