Abstract:
Galaxy surveys of the next decade will observe hundreds of millions of galaxies over unprecedented cosmic volumes and produce detailed 3D maps of galaxies. These maps encode the growth and expansion histories of the Universe that can be used to precisely test the standard “Lambda-CDM” cosmological model and probe the nature of dark energy. While, current analyses extract some of this cosmological information by summarizing the galaxy maps into 2-point clustering statistics, much more information still remain in the data. In my talk, I will present how we can use simulation-based inference and leverage high-fidelity cosmological simulations to extract the full cosmological information of galaxy surveys. Specifically, I will present SimBIG, a galaxy clustering analysis framework using simulation-based inference with normalizing flows. I will show the latest results from applying SimBIG to data from current galaxy surveys and showcase the improvements we find over the current baseline analyses. Lastly, I will discuss how SimBIG will be extended to the next-generation galaxy surveys to produce even more precise tests of the Lambda-CDM model and probe dark energy across cosmic history.
Bio:
ChangHoon Hahn is an assistant professor at the University of Arizona and Steward Observatory, where he is a member of the Arizona Cosmology Lab. His research focuses on applying astrostatistics and machine learning methods to the millions of galaxies observed by galaxy surveys to answer fundamental questions in cosmology and galaxy evolution. Before Arizona, he was a research scholar at Princeton and a postdoctoral fellow at Lawrence Berkeley National Lab and UC Berkeley Center for Cosmological Physics. He received his PhD from NYU Center for Cosmology and Particle Physics.
Summary:
Focus: Analysis of cosmological survey data
LambdaCDM model:
History of the universe:
Big Bang
Cosmic Inflation
Galaxies form, dark matter is 85% of matter
10B years: dominated by dark energy driving cosmological expansion
Model Parameters
How much matter
How much baryons
How fast universe is expanding
How ripples in early universe were distributed
How clumpy the universe is at large scales
Challenge: different families of measurements (early vs late universe) predict different values of of the expansion rate parameters
Using data from galaxy observations
Large-scale galaxy structures depend on
Expansion of universe
Gravitational attraction and motion of galaxy clusters
Sound wave propagation through early universe baryonic matter
Currently using coarse statistical analysis based on co-occurence of galaxy pairs/structures under different evolution/statistical assumptions
Approach:
Use simulations of galaxy to predict spatial distribution of galaxies and compare to galaxy survey data. Infer best simulation parameters based on simulation’s prediction error.
Quijote: n-body galaxy simulation
Molino: galaxy catalogs from Quijote
Insight: many different metrics of the distribution of galaxies based on regional structure; valuable area for exploration to improve model accuracy
Challenge: difficult to infer the fine-scale structure of the universe due to modeling and experimental challenges
Approach: generative model of universe structure
Run simulation many times to generate many possible galaxy structures
Compare predicted statistical distribution to real cosmological survey data
Use KL Divergence to differentiate predicted probability distribution from the real distribution
Use this to infer a tighter distribution of values of Lambda-CDM parameters
Running simulations is too expensive, so training a neural surrogate based on normalizing flows model
SimBIG: Simulation-based inference of galaxies
https://changhoonhahn.github.io/simbig/current/
Generate training data of galaxy observations
Train normalizing flow
Compare model’s predictions to real data
Use error to update model’s posterior
Survey: SDSS-III (https://www.sdss3.org/)
Used SimBIG to tighten estimates of key parameters, especially when using higher-order galaxy structure statistics based on relative locations of multiple nearby galaxies
1.9x tighter Sg parameter
1.5x tighter H0 parameter
Equivalent to collecting 4x more observational data
Future: additional data from galaxy surveys
Every decade we observe ~10x more galaxies
Dark Energy Spectroscopic Instrument (DESI): 4m Mayall telescope, 10b years of cosmic history
SuMIRe Prime Focus Spectrograph: 8.2m Subaru telescope, 12b years of cosmic history