Fast, high-fidelity Lyman $α$ forests with convolutional neural networks

Abstract

Full-physics cosmological simulations are powerful tools for studying the formation and evolution of structure in the universe but require extreme computational resources. Here, we train a convolutional neural network to use a cheaper N-body-only simulation to reconstruct the baryon hydrodynamic variables (density, temperature, and velocity) on scales relevant to the Lyman-α (Lyα) forest, using data from Nyx simulations. We show that our method enables rapid estimation of these fields at a resolution of ∼20kpc, and captures the statistics of the Lyα forest with much greater accuracy than existing approximations. Because our model is fully-convolutional, we can train on smaller simulation boxes and deploy on much larger ones, enabling substantial computational savings. Furthermore, as our method produces an approximation for the hydrodynamic fields instead of Lyα flux directly, it is not limited to a particular choice of ionizing background or mean transmitted flux.

LyMAS reloaded: improving the predictions of the large-scale Lyman-α forest statistics from dark matter density and velocity fields

“…Recently, Harrington et al (2021) have trained a convolutional neural network from hydrodynamical simulations of side 20 Mpc/h to predict both the density, the temperature and the velocity fields. This method is quite flexible and the predictions of the flux PDF and 1d power spectrum (i.e.…”

“…Both methods have proved to be again more accurate than the FGPA approach (which strongly relies on the DM smoothing scale) when reproducing line-of-sight observables, such as the PDF and power spectrum as well as the 3d flux power spectrum (5-20%). Finally, Machine Learning based methods start to be considered and lead to promising results (Harrington et al 2021;Sinigaglia et al 2021;Chopitan et al 2021).…”

The CAMELS Multifield Dataset: Learning the Universe's Fundamental Parameters with Artificial Intelligence

“…In cosmology, deep learning is being used to carry out many different complex tasks where traditional methods were slow, inaccurate or simply nonexistent. Examples of such tasks are speeding up simulations (He et al 2019;Alves de Oliveira et al 2020), obtaining super-resolution simulations (Kodi Ramanah et al 2020;Li et al 2021;Ni et al 2021), cleaning up astrophysics (Villanueva-Domingo & Villaescusa-Navarro 2020; Makinen et al 2021), painting stars and gas properties on the dark matter field (Wadekar et al 2021;Yip et al 2019;Zhang et al 2019;Kasmanoff et al 2020;Thiele et al 2020;Moews et al 2021;Jo & Kim 2019;Modi et al 2018;Tröster et al 2019;Harrington et al 2021), changing the cosmology of a simulation (Giusarma et al 2019), generating new data with desired properties (Dai & Seljak 2021;Böhm & Seljak 2020), and detecting anomalies (Storey-Fisher et al 2021), among many more 1 .…”

From EMBER to FIRE: predicting high resolution baryon fields from dark matter simulations with Deep Learning

“…Another interesting application is to learn the mapping between two matter components. A prominent example is to link the distribution of dark matter to specific baryonic fields like galaxies (Agarwal et al 2018;Zhang et al 2019;Jo & Kim 2019;Moster et al 2020), neutrinos (Giusarma et al 2019) as well as various gas fields (Tröster et al 2019;Zamudio-Fernandez et al 2019;Dai & Seljak 2020;Thiele et al 2020;Wadekar et al 2020;Lovell et al 2021;Harrington et al 2021;Prelogović et al 2021).…”