IllustrisTNG: a Universe in a Box

Dylan Nelson, Heidelberg University

Video Recording

Abstract:

Recently it has become possible to numerically simulate large, representative volumes of the Universe. These cosmological (magneto)hydrodynamical simulations solve for the coupled evolution of gas, dark matter, stars, and supermassive black holes interacting via the coupled equations of self-gravity and fluid dynamics, all within the context of an expanding spacetime.


The IllustrisTNG simulations are the current state-of-the-art in this context. They simultaneously resolve tens of thousands, to millions, of individual galaxies - with properties and characteristics in broad agreement with observational data of real galaxy populations. This enables many theoretical studies on galaxy formation and evolution, as well as large-scale structure and cosmology.


These calculations are challenging. Our largest runs to date evolve ~1e10 resolution elements for >1e7 timesteps, requiring ~100M core hours, parallelized on up to ~100k CPU cores, and typically run on Top10 HPC systems. Our numerical backbone is the AREPO "moving-mesh" code, which adopts a finite-volume type approach to solve the equations of ideal MHD on an unstructured, spatially and temporally adaptive discretization of space provided by a Voronoi tessellation. The enormous dynamic range of the problem, in both space and time, requires the use of a hierarchical, individual time stepping scheme. Simultaneously, key astrophysical phenomena always occur -below- the resolution scale, requiring extensive "sub-grid" models to capture processes such as radiative cooling, star formation, supernovae explosions, black hole dynamics, growth, and feedback energy.


Bio:

Dylan Nelson is currently an Emmy Noether Research Group Leader at the Institute for Theoretical Astrophysics, within the Center for Astrophysics (ITA/ZAH) of Heidelberg University, Germany. Previously he was a postdoctoral fellow at the Max Planck Institute for Astrophysics (MPA) in Munich, Germany. Five years ago he completed his PhD in astrophysics at the Center for Astronomy (CfA) at Harvard University, working with Lars Hernquist. His interests include theoretical modeling of cosmological gas accretion, the circumgalactic medium, the baryon cycle, and energetic feedback processes, particularly in their connections to the formation and evolution of galaxies and galactic structure over cosmic time.

Summary:

IllustrisTNG (https://www.tng-project.org/) models universe from scale of star clusters to galaxy clusters and 

- Incorporates different types of physics

   - Gravity (inter-galaxies, matter/dark matter)

   - Gas fluid dynamics

   - Magnetism

   - 1e6 range of spatial scales between smallest/largest structures

   - Large scales (inter-galactic) determine the matter flowing into/out of a galaxy

   - Smallest scales affect the larger ones (e.g. central black hole of a galaxy affects the galaxy's dark matter halo) 

 - Accounts for formation of galaxies and stars over cosmic time and captures both the evolution galaxy shapes (black holes, stars) and interactions between galaxies.

 - IllustrisTNG incorporates many numerical techniques to makes solving the equations tractable

   - Gravitational interactions modeled using hierarchical interaction trees

   - Magnetohydrodynamics: moving mesh defined via a Voronoi tessellation

- The finest-granularity they wish to model is still 1e6x smaller than the smallest scale IllustrisTNG can model

   - Processes below their resolution limit are modeled approximately via "sub-grid models"

   - e.g. individual stars are too small for IllustrisTNG

      - One "star particle" in IllustrisTNG is ~1e6 stars

      - These "star particles" do roughly the same things as real stars, with coarse adjustments for the missing physics within the particle

      - Sub-grid model parameterized from other simulations, observations, tuning parameters to maximize accuracy 

         - (e.g. what fraction of solar mass is radiated away via light)

   - IllustrisTNG has 10-50 free sub-grid parameters

      - Currently tuned by hand/heuristically

      - Wish to use a more statistically rigorous evaluation 

      - There's an opportunity for using ML approximations for the simulations to help tune the parameters (parameters that are best for the ML approximation are a decent guess for the best parameters for the full simulation)

      

Videos: