Simulating Cosmic Structure Formation with Neural Networks

Speaker: Drew Jamieson (MPIA) In Person and zoom

Zoom info: https://stanford.zoom.us/my/sihanyuan?pwd=QnpsUHZWWGJ2ekVYWmZVL3BmM0gzZz09

The next generation of galaxy surveys will produce vast amounts of high-quality data, helping us understand the Universe’s contents, history, and structure—and potentially uncover new physics. To make the most of this data, we need fast and accurate simulations of how cosmic structures—like galaxies and clusters—form over time. While N-body simulations are extremely precise, they are too slow for direct use in data analysis. In this talk, I will introduce a new method that uses a neural network to model cosmic structure formation. This model begins with the simple, linear initial conditions of the early universe and predicts how they evolve into the more complex, nonlinear structures seen in observations and simulations. The neural network is enhanced to consider the underlying cosmology and the Universe’s time evolution. A key feature of the model is that it enforces velocities equal time derivatives of displacements, enforcing a physical constraint that significantly improves accuracy. Training the network on a wide range of cosmic simulations, we achieve highly accurate predictions, even on the small scales where nonlinearities dominate. Compared with detailed N-body simulations, the model successfully captures the history of structure formation, including mergers trees of dark matter halos. This method is efficient enough for practical applications, such as generating mock galaxy catalogs and helping us infer the Universe's initial conditions and fundamental parameters.