Finding Universal Relations in Subhalo Properties with Artificial Intelligence

Helen Shao, Francisco Villaescusa-Navarro, Shy Genel, David N. Spergel, Daniel Anglés‐Alcázar, Lars Hernquist, Romeel Davé, Desika Narayanan, Gabriella Contardo, Mark VogelsbergerThe Astrophysical Journal 2022 · DOI

TL;DR

This paper uses machine learning to discover a universal relationship between the total mass of dark matter subhalos and their internal properties, achieving over 99% accuracy in predictions. This discovered relation, which appears to be linked to the virial theorem, holds across various cosmological simulations and even at different redshifts, suggesting a fundamental physical law at play.

What Problem Does This Paper Address?

Understanding the formation and evolution of galaxies is a central challenge in astrophysics. Galaxies are embedded within massive dark matter halos, and their properties are intricately linked to the properties of these halos and their substructures, known as subhalos. While many correlations between galaxy and subhalo properties are known, finding new, fundamental relationships in the high-dimensional space of these properties is difficult. Such relationships could provide crucial insights into the underlying physics governing galaxy formation. This paper aims to uncover such a universal relationship, specifically focusing on predicting the total mass of a subhalo from its other internal characteristics.

Methods

The authors employ a two-pronged approach:

  1. Neural Networks (NNs): They train deep neural networks to learn a function that maps various subhalo properties (like radius, velocity dispersion, gas metallicity, star formation rate, etc.) to the subhalo's total mass. This is a standard application of NNs for regression tasks. Key to their success is using a generic formalism that can search for relations in high-dimensional spaces and testing the network's ability to extrapolate to unseen simulations and redshifts. They also use saliency maps to identify the most important input features for the NN.
  2. Symbolic Regression: To interpret the complex relationships learned by the NNs, they use symbolic regression techniques (genetic programming and template fitting) to derive approximate analytical equations. This allows for a more physically interpretable representation of the discovered relation.

Datasets

The study utilizes data from several large-scale cosmological simulations:

  • IllustrisTNG100 and IllustrisTNG300: State-of-the-art magnetohydrodynamic simulations providing detailed galaxy formation physics.
  • CAMELS (Cosmology and Astrophysics with MachinE Learning Simulations): A suite of 1000 hydrodynamic simulations (CAMELS-IllustrisTNG and CAMELS-SIMBA) with varying cosmological parameters and astrophysical models, and 2000 N-body simulations (CAMELS-Nbody) for gravity-only comparisons.

These datasets provide a diverse range of subhalos with varying properties, cosmologies, and physics implementations, crucial for testing the universality of the discovered relation.

Key Insights and Ideas

  • A Universal Subhalo Mass Relation: The most striking finding is the discovery of a remarkably accurate and universal relation between subhalo total mass and its internal properties, holding across vastly different simulation setups.
  • Virial Theorem Connection: The discovered relation shows strong similarities to the virial theorem, suggesting that gravity and the virial equilibrium play a fundamental role in setting this relationship.
  • Importance of Proper Radius: The neural network's ability to extrapolate to higher redshifts is critically dependent on using the proper radius of the subhalo, not its comoving radius. This highlights the importance of physical quantities over cosmological ones for understanding fundamental relations.
  • Robustness to Simulation Differences: The learned relation is robust to variations in simulation resolution, cosmology, and even the underlying hydrodynamic solver and subgrid physics models, underscoring its fundamental nature.
  • Extrapolation Power: The neural networks and derived analytic equations demonstrate impressive extrapolation capabilities, predicting subhalo masses accurately even for systems and conditions not encountered during training.

Key Results

  • High Prediction Accuracy: Neural networks predict subhalo total mass with over 99% of predictions within 0.2 dex of the true value for the training simulations.
  • Extrapolation Success: The trained models maintain high accuracy (often >95%) when tested on subhalos from different simulations, different resolutions, and even at redshifts up to z=2.
  • Analytic Approximations: Derived analytic equations (particularly Eq. 8 and G1) approximate the NN relation, with some regimes where they are more accurate than the NNs themselves.
  • Feature Importance: The most important properties for predicting subhalo mass are identified as radius (R), velocity dispersion (σ), maximum circular velocity (Vmax), gas mass (Mg), and star formation rate (SFR).

Limitations and Future Work

  • Analytic vs. NN Accuracy: While the analytic equations capture the main trend, they generally do not reach the same level of accuracy as the neural networks, suggesting the NN has learned a more complex relationship.
  • Subhalo Definition Dependence: The authors acknowledge that the specific definition of subhalos and their properties (e.g., radius, velocity dispersion) used by the simulation's analysis pipeline could influence the results. Retraining and refitting might be necessary if different definitions are used.
  • Physical Interpretation: While the relation shows similarities to the virial theorem, a complete physical interpretation of how all the input variables contribute to this relation is still an area for future work.
  • Simpler Analytic Forms: The authors believe simpler and more accurate analytic equations can be found, leaving this as a direction for future research.
  • Observational Comparisons: The authors plan to extend this formalism to search for similar relations in observational data and derive analytic equations from them.

Summary updated

Abstract

Abstract We use a generic formalism designed to search for relations in high-dimensional spaces to determine if the total mass of a subhalo can be predicted from other internal properties such as velocity dispersion, radius, or star formation rate. We train neural networks using data from the Cosmology and Astrophysics with MachinE Learning Simulations project and show that the model can predict the total mass of a subhalo with high accuracy: more than 99% of the subhalos have a predicted mass within 0.2 dex of their true value. The networks exhibit surprising extrapolation properties, being able to accurately predict the total mass of any type of subhalo containing any kind of galaxy at any redshift from simulations with different cosmologies, astrophysics models, subgrid physics, volumes, and resolutions, indicating that the network may have found a universal relation. We then use different methods to find equations that approximate the relation found by the networks and derive new analytic expressions that predict the total mass of a subhalo from its radius, velocity dispersion, and maximum circular velocity. We show that in some regimes, the analytic expressions are more accurate than the neural networks. The relation found by the neural network and approximated by the analytic equation bear similarities to the virial theorem.
BibTeX
@article{shao2022finding,
  author = {Helen Shao and Francisco Villaescusa-Navarro and Shy Genel and David N. Spergel and Daniel Anglés‐Alcázar and Lars Hernquist and Romeel Davé and Desika Narayanan and Gabriella Contardo and Mark Vogelsberger},
  title = {{Finding Universal Relations in Subhalo Properties with Artificial Intelligence}},
  journal = {The Astrophysical Journal},
  year = {2022},
  doi = {10.3847/1538-4357/ac4d30},
}
Figure