One of the most fundamental questions in the studying the Large Scale Structure of the Universe is the effect of environment on the Galaxy formation and evolution. In addition to the spatial properties of galaxies, some of their characteristics such as star formation rate, metalicity, shape, and spin might depend on the environmental conditions, which is the dark matter distribution in this context.
Due to complicated physics involved in the galaxy formation, cosmologists usually run simulations to study these processes, which is time consuming and computationally expensive. In the past few years machine learning algorithms emerge as a useful assistant for theoretical and observational cosmologists.
In this project we propose to use deep learning method to establish a mapping between Galaxies in Hydrodynamic/N-body simulations and its underlying dark matter distribution, which may cancel the need for running costly simulations each time after changing the cosmological model.
The first aim of our project is to re-derive the results of a very recently published paper by Zhang et al. . We divided this project into sections which are explained below:
-
acquiring Data for sub-boxes of Illustris simulation with high/mediate resolutions to study the performance of our algorithm in different resolutions.
-
Sampling for training set and testing set
-
Learning the algorithm in the sub-boxes with two resolutions
-
Test the algorithm in other sub-boxes of the simulation
-
Train and test the algorithm for a larger high resolution box
We will use the data from IllustrisTNG simulation, which is the next generation of Illustris with more precise astrophysical models/processes. The IllustrisTNG project is an ongoing series of large, cosmological magnetohydrodynamical simulations of galaxy formation, based on (\Lambda CDM) cosmology, that uses Planck 2015 results as model parameters. These simulations include both dark matter-only and radiative runs which makes them a suitable source for our project. Indeed, due to computational limitations, our main focus will be on low resolution runs of IllustrisTNG ((L_{box} \approx 51.7) Mpc) that includes (2 \times 2160^3) elements .
We will use Convolutional Neural Networks (CNN), which are useful when dealing with high volume data (such as image processing, etc). In order to prevent over-fitting, several convolution and pooling operations should be done to reduce number of parameters and make feature maps with extracting global features of the data.
Several architectures for CNN exist: LeNet, U-Net, ResNet, Inception, etc. The original paper adopted U-net. However we will decide which architecture we would use after downloading and analyzing a portion of the original data. We will use Keras for implementing our CNN.
The main goal of this project is to find out the relation between dark matter field and galaxy positions. Besides, we are wishful to answer the following questions: If this relation exists, what are the parameters? Can we predict galaxy mass distribution, and more ambitiously, color, metalicity, star formation rate, etc. from a given dark matter field? And if there is no such a relation, then what are the other features that can be used in order to predict statistical properties of galaxies?