Computational Modeling of Large-Scale Structure With Abacus

Abstract

Interpreting galaxy surveys in a cosmological context requires an accurate forward model of large-scale structure. N-body simulations are the state-of-the-art tool for this but are not without their challenges. For one, they are computationally expensive, demanding large allocations of supercomputer time; for another, they are only as accurate as their discrete particle representation of dark matter allows. In my dissertation, I address both of these challenges. Abacus is a code for cosmological N-body simulations based on an exact decomposition of the near-field and far-field force, making it exceptionally accurate and fast. Using one dual-GPU node, Abacus can solve a supercomputer-sized N-body problem many times faster than other codes while retaining orders-of-magnitude higher force accuracy. We present a full description of the Abacus code, including discussion of the combination of mathematical techniques, software implementation, and commodity computer hardware that makes this possible.

Abacus's accuracy has allowed us to identify and correct non-physical effects that arise from the discrete N-body representation of dark matter. Using modified initial conditions, we can suppress discreteness errors in the late-time matter power spectrum by an order of magnitude, including effects that bias the outcome of linear perturbation theory. We have released a suite of more than 150 across 40 cosmologies simulations called "Abacus Cosmos" based upon these results. Finally, we develop a technique for maximizing the reuse of N-body simulations by perturbatively changing the background cosmology of a simulation output. By focusing on high accuracy for small changes in cosmology, the "warped" results are indistinguishable from a full N-body realization in key analysis metrics like the galaxy power spectrum. Abacus will be transformative for executing simulations of the size and fidelity required for analysis of surveys such as DESI, Euclid, and WFIRST.