Emulator-Based Inference of Cosmological Subgrid Models

TL;DR

This work advances emulator-based inference for cosmological subgrid physics by coupling Gaussian-process emulators (via SEPIA) to the CRK-HACC hydro code, calibrating five subgrid parameters against multiple baryonic observables (GSMF, CGD, f_gas) in a two-phase design that uses both moderate and large volume simulations. The analysis reveals two distinct AGN kinetic-feedback regimes, illustrating tension between cluster gas fractions and inner gas densities, and demonstrates that larger-volume simulations yield more reliable CGD constraints. The study also extends emulation to additional statistics (P_sub/P_grav, BHMSM, CSFR) and validates the calibrated parameter sets against a larger Frontier-E-Small run, highlighting volume effects and residual biases. Overall, the framework enables robust, multi-observable calibration of complex subgrid physics, essential for producing reliable predictions for survey-scale hydrodynamic simulations and for understanding baryonic impacts on small and large-scale structure.

Abstract

The formation of structure in the Universe at large scales is dominated by gravity, with baryonic physics becoming significant at $\sim{\rm Mpc}$ scales. To capture the impact of baryonic physics, cosmological simulations must model gas dynamics and a host of relevant astrophysical processes. A recent extension of the Hardware/Hybrid Accelerated Cosmology Code (HACC) couples its gravity solver with a modern smoothed particle hydrodynamics method. This extension incorporates sub-resolution models for chemical enrichment, black hole and star formation, AGN kinetic and thermal feedback, supernova-driven feedback, galactic winds, and metal-line cooling. We present an inference framework based on high-fidelity emulators to aid in model calibration against observational targets, e.g., the galaxy stellar mass function, radial gas density profiles, and the cluster gas fraction. The emulators are trained on simulation suites comprising 64 boxes with side-length $128\,h^{-1}$Mpc and 16 boxes with side-length $256\,h^{-1}$Mpc with $2\times 512^3$ and $2\times 1024^3$ particles, respectively. Our analysis reveals two distinct AGN kinetic feedback modes -- a low-feedback mode yielding strong agreement with the observed radial gas density profiles of massive X-ray clusters, and a high-feedback mode providing a better fit to cluster gas fraction data, but systematically underestimating gas densities in inner regions.

Figures (9)

• Figure 1: Results for the sensitivity analysis for the three observables ($\rm{GSMF}$, $\mathrm{CGD}$, and $f_\mathrm{gas}$) using emulator predictions. The impact of five subgrid parameters is investigated. Within each panel, the variation of a summary statistic is shown when a single parameter is varied, while the rest are fixed to the center of the experimental design.
• Figure 2: Right panel: Bayesian posterior distribution for the five subgrid parameters obtained via MCMC sampling in Phase-1. The forward model involves three combinations of the $\rm{GSMF}$, $f_\mathrm{gas}$, and $\mathrm{CGD}$. The plots show the effect using $f_\mathrm{gas}$ targets as a part of the calibration, which results in high feedback and deteriorated inner profiles of the $\mathrm{CGD}$ in the clusters. Left panels show the emulator predictions (solid gray, red, and blue lines) along with the errors corresponding to 5th and 95th percentiles. Black dots show the observational datasets listed in \ref{['sec:obs']}. Faint gray lines show the ensemble of training summary statistics.
• Figure 3: The MCMC posterior distribution with likelihood function that includes $\rm{GSMF}$ and $\mathrm{CGD}$, but not $f_\mathrm{gas}$. The Phase-1 emulators are used in the likelihood calculation. Right panel: Comparison of the emulator predictions at the best-fit parameters (solid blue lines) against the observational data compiled in \ref{['sec:obs']} (black dots). Faint gray lines show the ensemble of training summary statistics. The emulation results show an agreement with observed $\mathrm{CGD}$ profiles, but the $f_\mathrm{gas}$ predictions are higher than the observational data.
• Figure 4: Phase-2 constraints on the kinetic feedback parameters. Left panel: Bayesian posterior distribution of two subgrid parameters obtained using MCMC sampling. The blue contour shows the posterior corresponding to just the $\mathrm{CGD}$ (from the larger box). The text box lists the best-fit kinetic feedback parameters. Right panels: Emulated $\mathrm{CGD}$ profiles (blue lines) corresponding to the best-fit parameters with the target observational datapoints (black dots).
• Figure 5: Results for the sensitivity analysis for the three observables not used in parameter calibration. The impact of five subgrid parameters is investigated. Within each panel, the variation of a summary statistic is shown when a single parameter is varied, while the rest are fixed to the center of the experimental design.