A deep learning model to emulate simulations of cosmic reionization. (arXiv:1905.06958v1 [astro-ph.CO])
<a href="http://arxiv.org/find/astro-ph/1/au:+Chardin_J/0/1/0/all/0/1">Jonathan Chardin</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Uhlrich_G/0/1/0/all/0/1">Gr&#xe9;goire Uhlrich</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Aubert_D/0/1/0/all/0/1">Dominique Aubert</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Deparis_N/0/1/0/all/0/1">Nicolas Deparis</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Gillet_N/0/1/0/all/0/1">Nicolas Gillet</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Ocvirk_P/0/1/0/all/0/1">Pierre Ocvirk</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Lewis_J/0/1/0/all/0/1">Joseph Lewis</a>

We present a deep learning model trained to emulate the radiative transfer
during the epoch of cosmological reionization. CRADLE (Cosmological
Reionization And Deep LEarning) is an autoencoder convolutional neural network
that uses two-dimensional maps of the star number density and the gas density
field at z=6 as inputs and that predicts 3D maps of the times of reionization
$mathrm{t_{reion}}$ as outputs. These predicted single fields are sufficient
to describe the global reionization history of the intergalactic medium in a
given simulation. We trained the model on a given simulation and tested the
predictions on another simulation with the same paramaters but with different
initial conditions. The model is successful at predicting $mathrm{t_{reion}}$
maps that are in good agreement with the test simulation. We used the power
spectrum of the $mathrm{t_{reion}}$ field as an indicator to validate our
model. We show that the network predicts large scales almost perfectly but is
somewhat less accurate at smaller scales. While the current model is already
well-suited to get average estimates about the reionization history, we expect
it can be further improved with larger samples for the training, better data
pre-processing and finer tuning of hyper-parameters. Emulators of this kind
could be systematically used to rapidly obtain the evolving HII regions
associated with hydro-only simulations and could be seen as precursors of fully
emulated physics solvers for future generations of simulations.

We present a deep learning model trained to emulate the radiative transfer
during the epoch of cosmological reionization. CRADLE (Cosmological
Reionization And Deep LEarning) is an autoencoder convolutional neural network
that uses two-dimensional maps of the star number density and the gas density
field at z=6 as inputs and that predicts 3D maps of the times of reionization
$mathrm{t_{reion}}$ as outputs. These predicted single fields are sufficient
to describe the global reionization history of the intergalactic medium in a
given simulation. We trained the model on a given simulation and tested the
predictions on another simulation with the same paramaters but with different
initial conditions. The model is successful at predicting $mathrm{t_{reion}}$
maps that are in good agreement with the test simulation. We used the power
spectrum of the $mathrm{t_{reion}}$ field as an indicator to validate our
model. We show that the network predicts large scales almost perfectly but is
somewhat less accurate at smaller scales. While the current model is already
well-suited to get average estimates about the reionization history, we expect
it can be further improved with larger samples for the training, better data
pre-processing and finer tuning of hyper-parameters. Emulators of this kind
could be systematically used to rapidly obtain the evolving HII regions
associated with hydro-only simulations and could be seen as precursors of fully
emulated physics solvers for future generations of simulations.

http://arxiv.org/icons/sfx.gif