Speaker
Description
Cosmological simulations are an essential part of understanding the large-scale structure of the universe and testing current cosmological theories. However, they can be computationally expensive. In particular, modelling the chemical evolution and gas cooling that occurs inside galaxies can take up significant amount of the computational budget. Solving the complex, coupled differential equations are typically handled by chemical evolution libraries such as Grackle, which iteratively evolve chemical abundances and temperature. In this talk I will show how we are utilizing advances in machine learning to accelerate these calculations and by learning approximations of complex physical models without explicitly solving the full set of equations we are able to speed up calculations whilst retaining accuracy. Specifically, I will present how we generate our data, the architecture of the model and ultimately its performance. Finally, I will show how this approach offers a method for reducing computational cost in cosmological simulations, ultimately making it more feasible to study a wider range of cosmological theories and to run larger simulations.
