Simple lessons from complex learning: what a neural network model learns about cosmic structure formation.

Area of Science:

• Cosmology
• Astrophysics
• Machine Learning

Background:

• Cosmological N-body simulations are crucial for understanding the universe's evolution.
• Predicting the full phase space evolution in the nonlinear regime is computationally challenging.
• Neural networks offer a potential solution for accelerating these complex simulations.

Purpose of the Study:

• To train a neural network model for predicting cosmological N-body simulation outcomes.
• To assess the model's accuracy and generalization capabilities.
• To compare the model's performance against existing simulation methods.

Main Methods:

• Training a neural network on cosmological N-body simulation data.
• Testing the model on simple, well-understood initial conditions (spherical, plane waves).
• Evaluating predictions for density, displacement, and momentum power spectra against N-body and COLA results.

Main Results:

• The neural network accurately approximates the Green's function expansion in the nonlinear regime.
• The model generalizes well to simple initial conditions, indicating learned physical principles.
• The model achieves percent-level accuracy at nonlinear scales, outperforming COLA.

Conclusions:

• Neural networks can effectively learn and predict complex cosmological phenomena.
• The model's failures on orthogonal modes provide insights into its limitations and learning process.
• This approach offers a significant speedup and accuracy improvement for cosmological simulations.