Learning effective physical laws for generating cosmological hydrodynamics with Lagrangian deep learning
Abstract
Hydrodynamical simulations are crucial for accurately predicting cosmological observations, but they are computationally too expensive to simulate in a full survey volume. Gravityonly Nbody simulations reduce the computational cost but cannot directly model the observables such as stars and gas. Here we propose a deeplearning approach to generate these observables. The model consists of several layers of Lagrangian displacement field moving the particles. It requires only a handful of parameters to learn, in contrast to neural networks, and can be viewed as learning the effective physical laws. By combining Nbody solvers with our model, we are able to generate higheraccuracy hydrodynamic observables with computational costs several orders of magnitude cheaper than the traditional hydrodynamical simulations at the same resolution.
 Publication:

Proceedings of the National Academy of Science
 Pub Date:
 April 2021
 DOI:
 10.1073/pnas.2020324118
 arXiv:
 arXiv:2010.02926
 Bibcode:
 2021PNAS..11820324D
 Keywords:

 deep learning;
 Lagrangian approach;
 cosmological hydrodynamical simulation;
 Astrophysics  Cosmology and Nongalactic Astrophysics;
 Computer Science  Machine Learning
 EPrint:
 10 pages, 6 figures