Hierarchical Initial Condition Generator for Cosmic Structure Using Normalizing Flows

Sammanfattning

In this report, we present a novel Bayesian inference framework to reconstruct the three-dimensional initial conditions of cosmic structure formation from data. To achieve this goal, we leverage deep learning technologies to create a generative model of cosmic initial conditions paired with a fast machine learning surrogate model emulating the complex gravitational structure formation. According to the cosmological paradigm, all observable structures were formed from tiny primordial quantum fluctuations generated during the early stages of the Universe. As time passed, these seed fluctuations grew via gravitational aggregation to form the presently observed cosmic web traced by galaxies. For this reason, the specific shape of a configuration of the observed galaxy distribution retains a memory of its initial conditions and the physical processes that shaped it. To recover this information, we develop a novel machine learning approach that leverages the hierarchical nature of structure formation. We demonstrate our method in a mock analysis and find that we can recover the initial conditions with high accuracy, showing the potential of our model.