Oct 3 2019
History of the reionisation of the Universe in a cosmological simulation (top) and prediction with the artificial intelligence (bottom). Both maps show the ionization state of the hydrogen gas at different location in space as a function of the cosmological time (in billions of year since the Big-Bang). Yellow zones are ionized while purple pixels represents non ionized regions (neutral hydrogen gas). Both reionization histories are extremely similar. Therefore, the artificial intelligence has learnt the radiative transfer physics associated with the first generation of galaxies.
Evolution of the ionized fraction as a function of cosmological time in the simulation and the corresponding prediction with the artificial intelligence. Both curve are almost superimposed.
For the first time, a research team of the Astronomical Observatory of Strasbourg has trained an artificial intelligence to simulate the light of the first galaxies of the Universe. In a paper coming soon in the journal “Monthly Notices of the Royal Astronomical Society”, the research team has shown that convolutional neural networks are able to learn the physics of the interaction between light and matter in the surroundings of these galaxies. Such a technique allows to emulate simulation of this phenomenom without solving the complex system of coupled differential equations that usually have to be solved inside simulation codes that aim to study this phenomenon. Such an artificial intelligence model is about 100 times faster than usual simulations and produces a simulation emulation in a few minutes only.
More precisely, the model is designed to simulate the reionisation epoch of the Universe. This period is the last
major transition that the hydrogen gas in the cosmos experienced in the Universe history between the Big-Bang
and today. This transition took place during the first billion years of the Universe history about 13 billions years
ago. This period is the consequence of the emission of high energy photons from the first star in the first generation of galaxies that ionize and heat the hydrogen gas around them. This will create ionized “bubbles” in expansion around these sources of radiation up to the moment when these different ionized regions merge together, eventually leaving a complete ionized Universe about one billion years after the Big-Bang.
Understanding the course of this period is of prime importance for the scientific community in order to explain
the mechanisms of formation and evolution of galaxies. Having a detailed explanation of this process will make it
possible to to explain the morphology and the number of stars observed in galaxies such as our own Milky Way
or our neighbor the Andromeda galaxy. Directly observing this transition is very complicated because we have to
look at very early epochs which corresponds to observe very far in the cosmos. The upcoming new generation of
spatial and ground base telescopes promise to observe the reionisation epoch by 2025-2030. In the meantime, the community is already modelling the phenomenon thanks to numerical simulations with the aim to theoretically interpret the future observations.
Simulating the reionisation epoch is very complicated because of the large range of spatial scales to cover and
because of the speed of light which is a constraint to follow the radiation propagation. Thus, the last generation
of simulations requires millions of hours of calculation and using the largest supercomputers in the world. One of
the goals of the simulation community is to find a way to accelerate these calculations. Many suggestions already
exist but they all share the problem of sacrificing the precision of the results compared to real simulations.
In this context, the Strasbourg team had the idea of using already-run simulations of the reionisation epoch to train an artificial intelligence to understand the physics at play in these models. This artificial intelligence predicts the ionization map of the gas at every instant during the reionisation epoch. Such a technique reproduces the results from real simulations with high precision. This study is the starting point towards a new generation of simulation codes in astrophysics based on artificial intelligence. Such tools will allow the scientific community to run their own simulations much faster than what is possible currently. This will allow us to study different theoretical scenarii concerning the course of the reionisation epoch before the venue of the future observations.
Article reference : A deep learning model to emulate simulations of cosmic reionization. Jonathan Chardin, Grégoire Uhlrich, Dominique Aubert, Nicolas Deparis, Nicolas Gillet, Pierre Ocvirk, Joseph Lewis accepted the 7
september 2019 by Monthly Notices of the Royal Astronomical Society
Contact : Jonathan Chardin, jonathan.chardin@astro.unistra.fr, Observatoire Astronomique de Strasbourg, 11 rue
de l’Université, 67 000, Strasbourg
- Files:
lightcone_english.jpg (130 ko)- QHII_vs_treion_english.jpg (29 ko)
