Goku-ELG Documentation

Under Development Built with GPflow arXiv Preprint Coming Soon

Goku-ELG is a percent-level accurate cosmological surrogate model for emission-line galaxies (ELGs), replacing expensive N-body simulations for modeling galaxy clustering.

Overview

The Problem

Modeling cosmological observations with high fidelity typically requires computationally expensive forward simulations of large-scale structure. Because these simulations are slow and costly, researchers often resort to simplified or approximate models, sacrificing physical accuracy and precision.

Our Solution

Using a Bayesian experimental design strategy, we carefully select a limited number of simulation runs within a 10-dimensional cosmological parameter space. We then train a multi-fidelity Gaussian Process (GP) surrogate on these simulation results to emulate the observed clustering signal.

This emulator achieves percent-level cross-validation accuracy, enabling fast and reliable inference through Markov Chain Monte Carlo (MCMC) sampling without the need for repeated, expensive N-body simulations.

Key Features

  • High Accuracy: Percent-level cross-validation accuracy for galaxy clustering statistics

  • Multi-fidelity GP: Advanced Gaussian Process modeling combining multiple simulation fidelities

  • 10D Parameter Space: Comprehensive coverage of cosmological parameters

  • Fast Inference: Enables rapid MCMC sampling without repeated N-body simulations

  • Built on GOKU Suite: Leverages the GOKU simulation suite for training

Interactive Demo

Try our interactive galaxy clustering demo: Goku-ELG Interactive Demo

Indices and tables