ASTROPHYSICS

An international team of scientists at the Heidelberg Institute for Theoretical Studies (HITS), MIT, Harvard University and the University of Cambridge has carried out the “Illustris Simulation” on the SuperMUC and CURIE supercomputers, and created the largest and most sophisticated computational model of cosmic structure formation thus far.

Galaxies are typically comprised of several hundred billion stars and display a variety of shapes and sizes. Their formation is one of the most involved and complex problems of astrophysics. The standard model of cosmology conjectures that the Universe is dominated by unknown forms of matter and energy. We do not have insights yet into the true physical nature of dark matter and dark energy, but their impact can be understood with the help of supercomputers. To this end, previous simulations of the cosmos generated a cosmic web of matter concentrations that showed a passing resemblance to the galaxy distribution. However, they were not able to create elliptical and spiral galaxies and follow the small-scale evolution of interstellar gas and stars, which are closely linked to each other. The ambitious “Illustris” project has taken a major step towards addressing this issue.

In the world’s largest hydrodynamic simulation of galaxy formation, 13 billion years of cosmic evolution were followed in a cube of 3500 million light-years across, beginning 12 million years after the Big Bang. During this time, the “primordial soup” consisting of hydrogen, helium gas and dark matter formed increasingly big clumps, which were pulled together by gravity. Finally, galactic stellar systems developed, whose growth is regulated by the complex interactions between radiation processes, hydrodynamic shock waves, turbulent flows, star formation, supernova explosions and the energy supplied through growing supermassive black holes. The Illustris team was able to calculate these physical processes with the AREPO code developed at the Heidelberg Institute for Theoretical Studies. AREPO is a “moving mesh code”, which does not partition the simulated universe with a fixed grid but rather uses a movable and deformable mesh, which allows a particularly accurate processing of the vastly different size and mass scales occurring in individual galaxies.

The simulation was possible thanks to the access to supercomputer resources at the GCS and through PRACE. The first half of the simulation was carried on the supercomputers CURIE in France as part of PRACE allocation RA0844. It was then finished on SuperMUC at LRZ in Germany, through a GCS large-scale project under allocation pr85je. The total CPU cost of the main simulation was about 19 million CPU hours, with a total memory requirement of over 25 TB RAM and between 8192 and well over 20000 cores used to evolve it forward in time. The main simulation of the project employed more than 18 billion particles and cells and bridged a dynamic range of more than one million per space dimension. The generated data volume of over 200 Terabyte represented a massive challenge for the project, but could still be handled thanks to the network infrastructure of PRACE and the DFN, and the capable mass storage systems of SuperMUC and CURIE. This rich data makes it possible to study the evolution of about 50,000 well-resolved galaxies in detail and to make theoretical predictions for cosmological structure formation with high accuracy.

The initial results of the simulation project were reported in an article in Nature (Vogelsberger et al. 2014). The calculation yielded for the first time a realistic mix of elliptical and spiral galaxies, an elusive goal in previous works. Moreover, the simulation explains the enrichment of heavy elements (so-called “metals”) in neutral hydrogen gas, and at the same reproduces in detail the observed clustering signal of galaxies around galaxy clusters. These successes can be regarded as a confirmation of the standard model of cosmology. At the same time, the model allows a variety of novel predictions that yet have to be tested with a comparison to observational data.