Emulating baryons from dark matter simulations

Abstract

The interplay of dark and baryonic matter across cosmic time is complex, highly non-linear, and a crucial probe of multiple astrophysical processes. Cosmological hydrodynamical simulations drive our understanding of this dynamic, but they are computationally expensive and suffer from a trade-off between volume and resolution. Mauro Bernardini and colleagues present a new deep-learning framework to address these difficulties by predicting high-resolution gas properties from simulated dark matter distributions at a fraction of the computational cost.

The framework, EMBER-2, achieves this mapping using conditional generative adversarial networks, and incorporates time dependence efficiently by modulating the base convolution kernels with the global redshift information. When trained on simulation data out to redshift z = 6 from the Feedback in Realistic Environments (FIRE) project, the software is able to accurately emulate the two-dimensional total gas surface density, radial velocity and temperature, and the H i surface density based on the projected surface density and radial velocity of dark matter. The emulated distributions globally conserve mass within 5% (total gas) and 10% (H i), with the largest errors resulting from the scarcity of training information in the regions of highest density.