Speaker
Description
We present TUNeS (Temporal UNet emulator for Structure formation), a fast neural emulator for cosmological structure formation across redshift. Starting from the initial particle distribution, TUNeS predicts the evolved matter density field with a two-stage architecture that combines particle-based large-scale evolution and grid-based nonlinear refinement. The framework is designed to be naturally extendable to larger volumes through stitching, enabling continuous structure generation over extended spatial regions. Trained on only a small number of N-body simulations, TUNeS achieves good accuracy in both the power spectrum and non-Gaussian statistics, while requiring only about 25 seconds to generate a 256^3 grids density field from initial conditions on a single GPU. These features make it a promising approach for fast mock production and other applications requiring many high-fidelity matter realizations.