Sultan tells us about his work training neural networks on the neutral hydrogen density fields in the CAMELS simulations.

He uses a process known as normalising flows to find a mapping between the non-linear, very non-Gaussian 2D projected density field and a different Gaussian field. Once this mapping is found, the idea is that one can do full statistics on the non-linear field, by sampling from the Gaussian one. The bold ambition is to use this process to reduce the need for running computationally expensive hydrodynamical simulations – making it more feasible to get precise cosmological constraints from future surveys.

It’s early days in the project, but the results are already promising as the model is able to generate density fields that match both the histogram of density values and the power spectrum (as well as the scatter in both functions) reasonably well.

Sultan: https://sultanier.wixsite.com/website

The Paper: https://arxiv.org/abs/2110.02983

CAMELS playlist: https://www.youtube.com/playlist?list=PLvy7h0l2rJHq03inVPqYnC3llKt0IwwLT

CAMELS intro: https://www.youtube.com/watch?v=6Vgc72a_VpY&t=0s