The latest calculations of the key parameters in the standard model of cosmology are significantly more accurate than earlier methods that used the same galaxy distribution data.
The standard model of the universe is built around just six key numbers. By utilizing a novel AI-driven approach, researchers from the Flatiron Institute and their collaborators have successfully revealed insights hidden within galaxy distribution to estimate the values of five cosmological parameters with remarkable accuracy.
The findings mark a considerable enhancement over values that were previously obtained. This new technique has cut the uncertainty regarding the distribution of matter in the universe to less than half when compared to traditional methods that used the same galaxy data. Moreover, the AI-driven technique correlates well with estimates of cosmological parameters derived from other observations, like the universe’s oldest light.
In a series of recent papers, including one published on August 21 in Nature Astronomy, the researchers introduced their method, known as Simulation-Based Inference of Galaxies (SimBIG).
According to Shirley Ho, a co-author and group leader at the Flatiron Institute’s Center for Computational Astrophysics (CCA) in New York City, obtaining tighter constraints on these parameters using the same data is essential for exploring topics ranging from dark matter composition to the characteristics of dark energy responsible for the universe’s expansion. This is particularly relevant with upcoming surveys of the cosmos expected to commence in the next few years.
“Each of these surveys costs hundreds of millions to billions of dollars,” Ho explains. “These surveys exist because we aim to better understand these cosmological parameters. When viewed pragmatically, each of these parameters could represent tens of millions of dollars. Therefore, it’s vital to conduct the best analysis possible to extract maximum knowledge from these surveys and further our understanding of the universe.”
The six cosmological parameters provide insights into the quantities of ordinary matter, dark matter, and dark energy within the universe, as well as conditions following the Big Bang, such as the universe’s early opacity during its cooling phase and the distribution of mass—whether it is dispersed or clustered. According to Liam Parker, a co-author of the Nature Astronomy study and research analyst at CCA, these parameters are essentially the ‘settings’ that govern the universe’s behavior at large scales.
A primary method for cosmologists to calculate these parameters involves analyzing how galaxies cluster in the universe. Historically, these analyses focused solely on the broad distribution of galaxies.
“We haven’t been able to analyze the smaller scales,” states ChangHoon Hahn, the study’s lead author and an associate research scholar at Princeton University. “For the past few years, we’ve known there is additional information present; we just lacked a proper method to extract it.”
Hahn proposed a way to utilize AI to garner small-scale data. His strategy unfolded in two phases: initially, he and his colleagues trained an AI model to estimate cosmological parameters from simulated universes. Subsequently, they tested the model with actual galaxy distribution data.
In their training, Hahn, Ho, Parker, and their colleagues presented the model with 2,000 box-shaped simulations from the CCA’s Quijote simulation suite, where each universe was created using different cosmological parameter values. They even introduced imperfections typical of real galaxy survey data, influenced by atmospheric conditions and telescope characteristics, to provide the model with practical training. “That’s a significant number of simulations, but still manageable,” Hahn notes. “Without machine learning, hundreds of thousands would be necessary.”
Throughout this training, the model learned the relationship between cosmological parameter values and subtle variations in galaxy clustering, including the spacing between individual galaxies. SimBIG also adapted to gather information from larger clusters of galaxies by examining their arrangements, creating visual shapes like elongated triangles or compact equilateral triangles.
Once the model was trained, researchers tested it with 109,636 real galaxies recorded by the Baryon Oscillation Spectroscopic Survey. As anticipated, the model utilized both small-scale and large-scale details from the data to enhance the precision of the cosmological parameter estimates. The accuracy was so high that it equaled what a traditional analysis would achieve with about four times the number of galaxies. Ho emphasizes its importance, indicating that the universe has a finite number of galaxies. By improving precision with less data, SimBIG can expand the boundaries of what can be analyzed.
One thrilling potential application of this newfound precision, Hahn suggests, relates to the Hubble tension—a cosmological issue stemming from conflicting estimates of the Hubble constant, which measures the rate at which the universe is expanding.
Determining the Hubble constant involves assessing the universe’s size using ‘cosmic rulers.’ Estimates derived from the distance to supernovae in distant galaxies are roughly 10% higher than those based on variations in ancient cosmic light.
With new surveys set to come online over the upcoming years that will capture more of the universe’s history, combining data from these surveys with SimBIG will allow for a clearer understanding of the extent of the Hubble tension and whether it indicates a need for a revised universe model or can be reconciled. Hahn states, “If we can measure these quantities very accurately and confirm there is a tension, it may uncover new physics regarding dark energy and the universe’s expansion.”
Alongside Michael Eickenberg from the Flatiron Institute’s Center for Computational Mathematics (CCM), Pablo Lemos from the CCA, Chirag Modi from the CCA and CCM, Bruno Régaldo-Saint Blancard from CCM, Simons Foundation president David Spergel, Jiamin Hou from the University of Florida, Elena Massara from the University of Waterloo, and Azadeh Moradinezhad Dizgah from the University of Geneva, Hahn, Ho, and Parker contributed to the Nature Astronomy SimBIG study.
