Beschreibung
Cosmological simulations face a fundamental trade-off between volume and resolution: large boxes are representative of the real Universe, but have insufficient fidelity to model the detailed physics of galaxy formation, while high-resolution hydrodynamic simulations must be run in volumes far too small to be cosmologically representative. Zoom simulations can re-simulate small regions from a large parent box, capturing the rare regions not represented in small periodic boxes, but they cannot naively be used to measure clustering statistics at large distances. I will present a framework for combining many zoom regions together using neural density estimators to recover population statistics, conditioned on cosmology and astrophysics. Each region's measurement is conditioned on its known large-scale environment and cosmological parameters, and the global statistic follows by integrating the learned distribution over environments and over many parent boxes spanning different cosmologies. We demonstrate this with two emulators: the first predicts the halo mass function from individual zoom regions, and the second targets two-point clustering by measuring correlations between pairs of zoom regions and integrating these to recover the full galaxy 2PCF ξ(r). Both are validated on the Quijote BSQ suite, using sub-volumes as controlled proxy zooms. Finally, I will discuss a new zoom simulation suite in preparation designed to exploit this new technique.