Introduction
What is Goku-ELG?
Goku-ELG is a cosmological emulator for emission-line galaxies (ELGs), built using the GOKU simulation suite. It provides a fast and accurate surrogate model for predicting galaxy clustering statistics across a wide range of cosmological parameters.
The Science Behind It
For Astrophysicists
Resolving emission-line galaxies in N-body simulations requires high resolution, while achieving high-fidelity clustering statistics demands large volumes. We address both challenges efficiently by using machine learning models trained on the GOKU simulation suite (Yang et al. 2025).
Our method employs a multi-fidelity Gaussian Process model (Kennedy & O’Hagan 2000, Ho et al. 2021), along with elements of the Stochastic Variational GP framework, extended to support datasets with varying uncertainty levels through a modified likelihood.
Technical Approach
The emulator combines several advanced techniques:
Bayesian Experimental Design: Strategic selection of simulation runs in the parameter space
Multi-fidelity Gaussian Processes: Leveraging simulations at different resolutions and volumes
Modified Stochastic Variational Framework: Handling datasets with heterogeneous uncertainties
Halo Model Framework: Converting halo statistics to observable galaxy clustering
Key Summary Statistics
The emulator provides predictions for:
Halo Mass Function (HMF): \(\phi(M) dM\) - the abundance of dark matter halos as a function of mass
Halo-Halo Correlation Function: \(\xi_{hh}(r, M_{th1}, M_{th2})\) - spatial clustering of halos
Galaxy-Galaxy Correlation Function: \(\xi_{gg}(r)\) - observable galaxy clustering
Projected Correlation Function: \(w_p(r_p)\) - projected galaxy clustering
Computational Pipeline
The computational flow follows this sequence:
HMF Emulator → ϕ(M) dM
↓
Halo-Halo ξ Emulator → ξ_hh(M_th1, M_th2)
↓
Hankel Transform → P_hh(k, M_th1, M_th2)
↓
HOD Model → P_gg(k)
↓
Inverse Hankel Transform → ξ_gg(r)
Cosmological Parameters
The emulator covers a 10-dimensional parameter space:
\(\Omega_m\) - Total matter density
\(\Omega_b\) - Baryon density
\(h\) - Hubble parameter
\(A_s\) - Scalar amplitude
\(n_s\) - Scalar spectral index
\(w_0\) - Dark energy equation of state (present)
\(w_a\) - Dark energy equation of state (evolution)
\(N_{ur}\) - Effective number of ultra-relativistic species
\(\alpha_s\) - Running of spectral index
\(m_\nu\) - Neutrino mass
Performance Validation
Our emulator achieves:
< 1% accuracy in cross-validation tests for galaxy-galaxy clustering
Percent-level precision for halo mass function predictions
Robust extrapolation within the trained parameter space
Fast evaluation: Orders of magnitude faster than running N-body simulations
Applications
Goku-ELG is designed for:
Cosmological parameter inference from galaxy surveys
Mock catalog generation for survey planning
Fisher forecasts for upcoming surveys
Bayesian inference with MCMC or nested sampling
Rapid prototyping of galaxy clustering models
Target Surveys
This emulator is particularly suited for:
HETDEX (Hobby-Eberly Telescope Dark Energy Experiment)
DESI (Dark Energy Spectroscopic Instrument)
Euclid
Roman Space Telescope
Other emission-line galaxy surveys