PI: Baojiu Li
A landmark achievement in 2025 was the completion of the MillenniumTNG-XXL (MTNG-XXL) run, a flagship calculation of the Virgo Consortium. This simulation pushes the boundaries of large-volume cosmological simulations, evolving the universe across 13.7 billion years of cosmic history using 1.1 trillion dark matter particles (10240³) within a colossal (3 Gpc)³ volume. The simulations used 34,524 COSMA-8 cores and 64.5M cpu-hours in total.
Figure 1: Massive neutrinos (top) and cold dark matter (bottom) distributions on the past lightcone of a fiducial observer as seen in MTNG-XXL. As cosmic expansion slows down the neutrinos at late times, they start to cluster.
A key technical innovation is the application of the so-called method (Elbers et al. 2021) — developed by members of the Virgo UK — which efficiently captures the influence of massive neutrinos (represented by 2,560³ particles) on the late-time matter distribution. The underlying simulation code, a modified version of the publicly available GADGET-4, was upgraded to self-consistently compute the Hubble expansion rate in the presence of radiation components (photons and massless neutrinos) as well as massive neutrinos, carefully accounting for their gradual transition from the relativistic to the non-relativistic regime. This ensures the code exactly reproduces the linear evolution of cosmic structure even on the small scales where massive neutrinos introduce a characteristic scale dependence in the linear growth factor.
MTNG-XXL includes on-the-fly halo and subhalo finding with simultaneous merger tree construction, allowing the detailed assembly histories of structures to be captured without storing a prohibitive number of simulation snapshots – a saving of approximately 4.1 TB of long-term storage. MTNG-XXL also records six different particle lightcone configurations, enabling the construction of mock galaxy catalogues in observationally motivated geometries, and produces weak-lensing mass-shell maps at a resolution of 1.8 billion pixels. The scale of the outputs is extraordinary. The simulation contains ~1.16 billion dark matter haloes at redshift zero, 1.46 billion gravitationally bound subhaloes, and ~1 billion distinct merger trees linking some 150 billion subhaloes across cosmic time. These halo catalogues are then processed with the state-of-the-art L-GALAXIES semi-analytic model, which solves coupled differential equations representing the physical processes of galaxy formation along the merger trees, ultimately producing catalogues of around 1.5 billion galaxies.
The scientific results are equally compelling. Galaxy statistics derived from the MTNG-XXL catalogues (including abundances and clustering in both real and redshift space) show excellent agreement with observations from major surveys such as SDSS and DESI, validating both the realism of the galaxy formation model and the power of the simulation’s enormous volume. The Baryon Acoustic Oscillation feature, a primary target of modern redshift surveys, is detected with exceptional clarity over the full diversity of galaxy types. The simulations reveal the imprint of massive neutrinos on galaxy formation: neutrino mass suppresses the star-formation history of galaxies across cosmic time, reduces the galaxy stellar mass function relative to massless-neutrino models, and leaves a measurable signature in galaxy clustering statistics. This demonstrates the unique scientific capability that DiRAC’s memory-intensive computing resources make possible.