The impact of stochastic modeling on the predictive power of galaxy formation simulations

Abstract

All modern galaxy formation models employ stochastic elements in their sub-grid prescriptions to discretise continuous equations across the time domain. In this paper, we investigate how the stochastic nature of these models, notably star formation, black hole accretion, and their associated feedback, that act on small (< kpc) scales, can back-react on macroscopic galaxy properties (e.g. stellar mass and size) across long (> Gyr) timescales. We find that the scatter in scaling relations predicted by the EAGLE model implemented in the SWIFT code can be significantly impacted by random variability between re-simulations of the same object, even when galaxies are resolved by tens of thousands of particles. We then illustrate how re-simulations of the same object can be used to better understand the underlying model, by showing how correlations between galaxy stellar mass and black hole mass disappear at the highest black hole masses (M BH > 10 8 M ), indicating that the feedback cycle may be interrupted by external processes. We find that although properties that are collected cumulatively over many objects are relatively robust against random variability (e.g. the median of a scaling relation), the properties of individual galaxies (such as galaxy stellar mass) can vary by up to 25%, even far into the well-resolved regime, driven by bursty physics (black hole feedback) and mergers between galaxies. We suggest that studies of individual objects within cosmological simulations be treated with caution, and that any studies aiming to closely investigate such objects must account for random variability within their results.

Robust Field-level Likelihood-free Inference with Galaxies

“…These simulations were run with the parallel cosmological Tree-PM code P-Gadget3 (Springel 2005). The code uses an entropy-conserving formulation of smoothed particle hydrodynamics (SPH; Springel & Hernquist 2002), with SPH modifications according to Dolag et al (2004Dolag et al ( , 2005Dolag et al ( , 2006 2015), with some parameter changes (Borrow et al 2022). The full model will be described in J.…”

Satellites of Milky Way- and M31-like galaxies with TNG50: quenched fractions, gas content, and star formation histories

“…7 and 8 as well as in Joshi et al 2021), some non-negligible residual stochasticity seems to be inherent to galaxy formation and evolution (see e.g. Keller et al 2019;Genel et al 2019;Borrow et al 2022, for discussions on the butterfly effect).…”

Hydrodynamical Simulations of the Galaxy Population: Enduring Successes and Outstanding Challenges

“…Many subgrid models employ stochastic treatments of continuous processes, such as the conversion of gas particles or cells into stellar particles, or injecting energy kinetically and/or thermally into the gas through feedback processes. A number of recent studies have examined the impact of this subgrid stochasticity on emergent macroscopic properties, in simulations of individual galaxies (Keller et al 2019;Davies et al 2021Davies et al , 2022 and periodic volumes (Genel et al 2019, Borrow et al 2022b, finding that small differences due to stochasticity can propagate into significant differences in present-day properties, even for galaxies resolved with many thousands of particles. The systematic uncertainty due to stochasticity is more significant in smaller samples of galaxies, i.e., zooms and small periodic volumes, and needs to be considered when interpreting results, including when analyzing small simulations for the purposes of model calibration.…”