Deciphering the initial conditions and formation history of cosmic structure with deep learning
|Deciphering the initial conditions and formation history of cosmic structure with deep learning
|Ludvig Doeser <firstname.lastname@example.org>
|2024-01-03 – 2024-08-01
The dynamical evolution and state of the Universe are determined by two fundamental properties: 1) the initial conditions, and 2) the laws of physics. Although the standard cosmological model has successfully withstood rigorous observational scrutiny, the detailed mechanisms underlying the origin of cosmic structures as well as the accelerated expansion of our Universe remain elusive. The synergy between accurate and robust physical modeling and insights derived from next-generation astronomical surveys is therefore becoming increasingly pivotal.
As per our present understanding, the spatial distribution of cosmic matter as traced by galaxies originates from primordial quantum fluctuations. These seed fluctuations were generated during the early epoch of inflation in a hot "Big Bang" and have since then grown via gravitational amplification to form a giant cosmic web of dark matter, eventually aggregating into massive clusters and filamentary cosmic structures.
Our research group, led by Assoc. Prof. Jens Jasche, at the Physics Department (“Fysikum”) at Stockholm University, has pioneered innovative data analysis methodologies to leverage the physical information contained in the large-scale structure of the Universe. By reconstructing comprehensive 3D maps of dark matter density and velocity fields from observed galaxy clustering data, we enable the extraction of significant cosmological information often overlooked by current methodologies. By considering higher-order moments of the cosmic matter distribution, we provide novel insights into the growth of structure. The reconstructions of the spatial matter distribution provide insights that serve as a cornerstone for testing fundamental physics and exploring the dynamic evolution of the Universe. The detailed spatial configuration and dynamics of the cosmic matter distribution preserve a clear imprint of the initial conditions. Additionally, they encode the fascinating physical processes that have intricately shaped the cosmic landscape over the remarkable period of 13.8 billion years.
Unlocking the full potential of next-generation cosmological data, such as the Vera Rubin Observatory's Legacy Survey of Space and Time and the Euclid satellite, requires navigating the balance between sophisticated physics models and computational demands. The selection of a physical model must align with recently established accuracy requirements for the inference process, ensuring the model's capability to faithfully recover the initial conditions. Recently, we proposed a solution by incorporating machine learning-based field-level emulators within the "Bayesian Origin Reconstruction from Galaxies" (BORG) inference algorithm. The emulators achieve remarkable accuracy compared to highly sophisticated N-body solvers while significantly reducing evaluation time. Leveraging its differentiable neural network architecture, the emulator enables efficient sampling of the high-dimensional space of plausible cosmic initial conditions.
Expanding upon this research, our objective is now to reconstruct cosmic initial conditions within the spatial domain encompassing galaxies from the 2M++ catalog. Our study introduces a novel approach, utilizing a swift and precise deep learning-based neural network model to unveil the intricate dynamics of structure formation. This project holds particular relevance in light of upcoming cosmological surveys, where the anticipated abundance of data in the order of billions of galaxies necessitates accurate modeling to optimally extract physical information.