Cosmic inflation postulates that a brief burst of rapid expansion in the very early universe shaped the distribution of galaxies we see today, but the physics behind what caused inflation remains uncertain. Motivated by this question, scientists are studying how galaxies are spread across the cosmos to uncover subtle correlations that could reveal more about inflation and the physics of the early universe. To support this research, Argonne scientists used the ALCF’s Aurora supercomputer to complete two of the largest cosmological simulations ever performed to aid in the analysis of modern surveys, including the DOE-supported Dark Energy Spectroscopic Instrument (DESI) and NASA’s SPHEREx mission, both of which are currently collecting data.
Challenge
The simplest, well-motivated inflation theories predict that primordial fluctuations were Gaussian to very high precision, meaning that they followed the simplest possible form of randomness. Any departure, known as primordial non-Gaussianity, could point to new mechanisms and help distinguish among competing theories, but the signal is expected to be subtle. Capturing it requires massive simulations that represent the largest cosmic scales while maintaining enough resolution to model the dark matter halos that host galaxies, since the galaxy distribution is the final observable. Achieving both enormous volume and sufficient resolution in a single calculation is extremely challenging, possible only with exascale computing.
Approach
Using the HACC code on Aurora, the team carried out two gravity-only simulations that each evolved approximately 13.4 trillion dark matter particles in a cube 28.6 billion light-years across. Each simulation ran on 8,100 Aurora nodes using roughly 50,000 GPUs simultaneously. Aurora’s large memory capacity helped the team fit the particles needed to combine vast volume with sufficient resolution, while its DAOS storage system supported intensive data writing. Completed in roughly four days, the two simulations wrote more than 85 petabytes, primarily as restart checkpoints required to protect calculations at this scale, with approximately 8 petabytes retained for scientific analysis.
Results
The calculations are two of the largest individual cosmological simulations ever performed by particle count. The simulations contain the same cosmology and non-Gaussian signal, but their random initial phases are flipped. Acting as statistical mirror images, they allow much of the random variation to cancel when analyzed together, providing a cleaner measurement of the underlying physics. The signal strength was also blinded so survey teams can test whether their pipelines recover the hidden value before analyzing real data. The simulations will enable synthetic full-sky maps while resolving the dark matter halos that host galaxies observed by DESI, SPHEREx, and other surveys.
Impact
The team’s massive simulations provide a foundation for advancing our understanding of inflation and the early universe, enabling researchers to validate analysis methods, identify biases, and improve measurements from current and future cosmology surveys.
