Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Revealing Local Cosmic Web Galaxies

Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Draft version March 19, 2021

Typeset using LATEX twocolumn style in AASTeX63

Revealing the Local Cosmic Web from Galaxies


by Deep Learning
ᆼᄉ
Sungwook E. Hong (ᄒ
ᅩ ᆼᄋ
ᅥ ᆨ) ,1, 2 Donghui Jeong,3 Ho Seong Hwang,4, 2 and Juhan Kim5

1 Natural Science Research Institute, University of Seoul, 163 Seoulsiripdaero, Dongdaemun-gu, Seoul, 02504, Republic of Korea
2 Korea Astronomy and Space Science Institute, 776 Daedeokdae-ro, Yuseong-gu, Daejeon 34055, Republic of Korea
3 Department of Astronomy and Astrophysics, and Institute for Gravitation and the Cosmos, The Pennsylvania State University,

University Park, PA 16802, USA


4 Astronomy Program, Department of Physics and Astronomy, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826,

Republic of Korea
5 Center for Advanced Computation, Korea Institute for Advanced Study, 85 Heogiro, Dongdaemun-gu, Seoul, 02455, Republic of Korea

ABSTRACT
The 80% of the matter in the Universe is in the form of dark matter that comprises the skeleton
of the large-scale structure called the Cosmic Web. As the Cosmic Web dictates the motion of all
matter in galaxies and inter-galactic media through gravity, knowing the distribution of dark matter
is essential for studying the large-scale structure. However, the Cosmic Web’s detailed structure is
unknown because it is dominated by dark matter and warm-hot inter-galactic media, both of which
are hard to trace. Here we show that we can reconstruct the Cosmic Web from the galaxy distribution
using the convolutional-neural-network-based deep-learning algorithm. We find the mapping between
the position and velocity of galaxies and the Cosmic Web using the results of the state-of-the-art
cosmological galaxy simulations, Illustris-TNG. We confirm the mapping by applying it to the EAGLE
simulation. Finally, using the local galaxy sample from Cosmicflows-3, we find the dark-matter map
in the local Universe. We anticipate that the local dark-matter map will illuminate the studies of the
nature of dark matter and the formation and evolution of the Local Group. High-resolution simulations
and precise distance measurements to local galaxies will improve the accuracy of the dark-matter map.

Keywords: Local Group; dark matter; large-scale structure of universe

1. INTRODUCTION dance cosmological model. Accounting for the measured


Since Fritz Zwicky inferred its existence from the large expansion rate of the Universe (Planck Collaboration
velocity dispersion of the Coma cluster (Zwicky 1933) et al. 2018) requires the matter component whose en-
and Vera Rubin confirmed it with the flat rotation curve ergy density is over five times larger than that of atoms
of galaxies (Rubin & Ford 1970), astronomers have been for which the robust upper limit comes from big-bang
only strengthening the necessity of the non-baryonic nucleosynthesis (Cooke et al. 2014). The observed large-
matter providing excess gravity. We call that dark mat- scale distribution of galaxies (Anderson et al. 2014) and
ter. The most substantial pieces of evidence include the map of weak-gravitational lensing potential (Abbott
an excessive mass-to-light ratio in the dwarf galaxies et al. 2018) also require the dark matter providing the
(Aaronson 1983), the mismatch between the X-ray map skeleton of the large-scale structure within which clouds
(gas distribution) and the weak gravitational lensing of atoms collapse to form stars and galaxies (Davis et al.
map (mass distribution; Clowe et al. 2006), as well as the 1985).
disparity between the heights of even- and odd-acoustic With the essential role that dark matter plays in mod-
peaks in the temperature power spectrum of the cos- ern astronomy and cosmology, in the past few decades,
mic microwave background (Larson et al. 2011). Dark there have been continuous efforts to search for the na-
matter is also an indispensable component of the concor- ture of dark-matter particles in the particle accelera-
tors (Atlas Collaboration 2019; Vannerom 2019), cos-
mic rays (Giesen et al. 2015), gamma-rays (Ackermann
Corresponding author: Donghui Jeong et al. 2015), and high-energy neutrinos (Aartsen et al.
djeong@psu.edu 2018). Beyond the Milky-way halo, there have also
2 Hong et al.

been recent studies focusing on the dark-matter signals output to the observation. However, this observational
from the extra-galactic sources by cross-correlating the constraint for the fully evolved galaxy distribution is
high-energy cosmic rays with the distribution of galaxies non-trivial to implement because the simulation needs
(Fornasa et al. 2016; Fang et al. 2020) and dark matter the density distribution at the initial time. Alterna-
traced by weak-gravitational lensing (Tröster et al. 2017; tively, the Bayesian Origin Reconstruction from Galax-
Ammazzalorso et al. 2020). All searches for the dark ies (BORG; see, e.g., Jasche & Wandelt 2013; Jasche
matter particles thus far, however, have not concluded et al. 2015) approach uses the multiple Gaussian pro-
with a firm detection. They have been only narrowing cesses to draw the probability distribution of the ini-
down the possible dark-matter masses and the interac- tial density perturbation from a given galaxy distribu-
tion strengths among dark matter particles as well as tion. As based on the dark-matter density field evo-
between dark matter and atoms (Akerib et al. 2017; Ar- lution by second-order Lagrangian perturbation theory
cadi et al. 2018). For these efforts of searching for the (2LPT) and linear galaxy bias model, the method is also
nature of dark matter, the most basic information cur- limited to, again, the scale larger than a few Mpc where
rently lacking is the distribution of the dark matter, or the 2LPT and linear bias models are accurate.
Cosmic Web, in the local large-scale structure beyond Here we overcome the challenges by taking a novel
the Milky-way halo. Of course, we have a good reason approach based on deep learning (DL). DL, as well as
to believe that dark-matter halos surround each galaxy a conventional machine learning technique, has been in-
in the Universe. It is, however, also well known that the troduced to measure the dark matter distribution from
galaxies are biased, rather than faithful, tracers of the weak gravitational lensing or spatial distribution of dark
large-scale structure (Desjacques et al. 2018). matter halos (e.g., Modi et al. 2018; Shirasaki et al. 2019;
In this article, we shall present a novel method of un- Jeffrey et al. 2020). On the contrary, our DL approach
veiling the Cosmic Web in the local Universe. As dark aims to reconstruct the local dark-matter map down to
matter is dark, of course, we cannot observe them di- an Mpc-scale by incorporating all information in the ob-
rectly from the telescope. The only guaranteed way of served galaxy data: the spatial distribution and the ra-
searching for the dark matter is the same method for dial peculiar velocity of galaxies. We use the convolu-
their discovery, through their gravitational influence on tional neural network (CNN)-based DL algorithm to find
visible objects. On the inter-galactic scales, dark matter the mapping between the local dark-matter distribution
dominates the gravitational interaction and determines and the observed positions and the radial peculiar ve-
the cosmic velocity flow. We can, therefore, infer the locities of local galaxies.
distribution of dark matter by carefully studying the dis- The structure of this paper is as follows. In Section 2,
tribution and motion of galaxies. Taking the observed we describe the simulation and observational data used
distribution of galaxies and their peculiar velocity flow, for DL training and prediction, respectively. In Sec-
in what follows, we shall decipher the dark matter dis- tion 3, we will briefly describe our DL architecture and
tribution, or Cosmic Web, within local ∼ 20 Mpc/h. the evaluation of our DL model. In Section 4, we will
When reconstructing the local dark-matter distribu- show the reconstructed local dark matter map and its
tion directly from observed galaxy distributions, we face statistical robustness. We will summarize our result in
the following challenges. First, the local galaxy distribu- Section 5.
tion at the low Galactic latitudes is hidden behind the Throughout the paper, we assume a standard ΛCDM
intense radiation from the Galactic disk and contami- cosmology in concordance with the Planck 2018 anal-
nated by the interstellar gas and dust, which makes it ysis (Planck Collaboration et al. 2018): (Ω0m , Ω0Λ , h) =
hard to obtain the complete map of the galaxy distribu- (0.31, 0.69, 0.6777). It is similar to the standard cos-
tion. Second, even if we had the complete map of galax- mologies adopted in Illustris-TNG and EAGLE sim-
ies, they are biased tracers of the large-scale structure; ulations: (Ω0m , Ω0Λ , h) = (0.3089, 0.6911, 0.6774) and
that is, the distribution of galaxies does not necessarily (0.307, 0.693, 0.6777), respectively (Springel et al. 2018;
reflect the distribution of dark matter. Schaye et al. 2015).
Previous attempts (Gottloeber et al. 2010; Libeskind
et al. 2010; Carrick et al. 2015; Lavaux & Jasche 2016; 2. DATA
Carlesi et al. 2016) of making the local dark-matter 2.1. Observational Data: Cosmicflows-3
map, therefore, have relied on the cosmological sim-
We use the Cosmicflows-3 galaxy catalog (Tully et al.
ulations constrained by the smoothed density field at
2016, CF3 hereafter), one of the most comprehensive
high-Galactic latitudes. Typically, a smoothing scale of
galaxy catalogs that provide distance, radial peculiar ve-
a few Mpc is employed when matching the simulation
locity, and luminosity of 17,647 galaxies up to 200 Mpc.
Local Dark-Matter Map from Deep Learning 3

To produce a fair galaxy sample over the given re- after) and the difference of peculiar velocity between the
gion, we make the volume-limited sub-sample of the CF3 target galaxy and center galaxy.
as follows. First, as the number density of the CF3 For the low-resolution dark-matter map with V =
galaxies close to the Galactic plane (Galactic latitude (40 Mpc/h)3 , we use the TNG300-1 from the Illustris-
|b| < 10◦ ) is lower than average, we only use the galaxies TNG simulations, whose volume and number of par-
at |b| > 10◦ . Also, we use the B-band absolute magni- ticles are V = (205 Mpc/h)3 and 25003 , respectively
tude (MB ) compiled from Lyon Extragalactic Database (TNG300 hereafter). Note that the amplitude of the
(LEDA; Paturel et al. 2003) as a proxy of the stellar mass luminosity function of TNG300 is lower than the obser-
(M? ; Wilman & Erwin 2012). We set the B-band mag- vation and TNG100, mainly due to the lower spatial res-
nitude −15 as the selection criterion, which is sufficient olution of the simulation (Pillepich et al. 2018). There-
for covering the 20 Mpc/h- and 40 Mpc/h-cubic volume fore, we also apply the resolution correction to find the
around the Milky-way galaxy. We have also tested the center and target galaxies using the number density ob-
cases with MB < −16 and −17 and found no noticeable tained from TNG100 rather than directly using the face
difference of the predictions from the fiducial choice (see values of M? or MB . We also use the TNG300-1-Dark,
Section 4). Note that we have not used the KS -band ab- a dark-matter-only counterpart of the TNG300, to test
solute magnitude, one of the best-known tracers of the how baryonic physics affects our result. We select the
stellar mass (Bell et al. 2003) because that information center and target galaxies by finding the mass cut of
is missing for about 30% of the galaxies in our sample dark matter halos with the same number density. The
(Lavaux & Hudson 2011; Huchra et al. 2012). result from the TNG300-1-Dark is similar to or slightly
We calculate the radial peculiar velocity by subtract- worse than TNG300 (see Section 4).
ing the Hubble flow from the velocity in the Galactic Also, we use the RefL0100N1504, a reference simula-
Standard of Rest (VGSR ; Kourkchi et al. 2020). Note tion with V = (67.77 Mpc/h)3 and 15043 dark-matter
that we do not use the velocity in the cosmic mi- and gas particles from the EAGLE simulation suite
crowave background (CMB) standard of rest (VCMB )to (Schaye et al. 2015; Crain et al. 2015, EAGLE hereafter),
reduce any bias that might be introduced in the conver- to check the fidelity of our result. For the center galaxies,
sion. Instead, when generating training and test samples we use the same selection criterion to TNG100 and find
from simulation data, we include the peculiar motion of 478 center galaxies. For the target galaxies, however,
the Milky-way corresponding galaxy in each simulation. we do not directly use MB . It is because the luminos-
There exists a difference on the Hubble constant be- ity function of EAGLE is reliable only for bright galax-
tween recent CMB observations (H0 = 67.77 km/s/Mpc; ies (MB . −18) since the EAGLE simulations calculate
Planck Collaboration et al. 2018) and the best-fit from the luminosity only to massive galaxies (M? ≥ 108.5 M ;
the CF3 (H0 = 75 km/s/Mpc; Tully et al. 2016). In this Camps et al. 2018). Instead, similar to TNG300, we use
study, we have tested both values and find that the ef- the galaxy number density obtained from TNG100 to
fect from the different Hubble constants stays within the find the stellar mass cut of target galaxies.
uncertainty of the dark-matter map (see Section 4).
3. METHODS
2.2. Simulation Data: Illustris-TNG & EAGLE 3.1. Deep Learning Architecture
We use the TNG100-1, a simulation with a comoving We construct the deep learning architecture using con-
volume V = (75 Mpc/h)3 and 18203 dark-matter and volutional neural network (CNN) that highlights fea-
gas particles from the Illustris-TNG simulation suite tures in the data by a series of convolutions, resulting
(Springel et al. 2018; Pillepich et al. 2018; Marinacci in so-called hidden layers. By varying the convolution
et al. 2018; Naiman et al. 2018; Nelson et al. 2018, filters, one can extract different physical features in the
2019), as our high-resolution simulation data (TNG100 data. Specifically, we use a CNN architecture similar to
hereafter). To mimic the observation from the Milky- the U-Net (Ronneberger et al. 2015) or V-Net (Milletari
way galaxy, we select 988 galaxies with stellar mass et al. 2016) to predict the dark-matter density field from
4 × 1010 M < M? < 1011 M (center galaxies here- the galaxy position and radial peculiar velocity (see Fig-
after)by adopting that the Galactic stellar mass is about ure 1). Our CNN architecture consists of the following
5.2 × 1010 M (Licquia & Newman 2015). Around each two stages: the encoding stage (Input–ConvNs ) with in-
center galaxy, we make a sub-cube with 20 Mpc/h box- creasing number of filters and decreasing the size of hid-
size and calculate the dark-matter density field within den layers, and the decoding stage (UpConvNs –Output)
the 643 uniform grid. We also calculate the relative po- with decreasing number of filters and increasing the size
sition of galaxies with MB < −15 (target galaxies here- of hidden layers. Here, Ns denotes the spatial size of
4 Hong et al.

Figure 1. The convolutional neural network (CNN) architecture used for TNG300. We denote the layer size by the quadruple
where the spatial dimension (2n , 2n , 2n ) follows the number of channels. The size (except the number of filters) of each layer for
TNG100 is the half of TNG300.
hidden layers. To retain the small-scale spatial resolu- `0 ∈ [1, Nch,0 ]), to obtain the output C as
tion, we also attach the hidden layers in the equivalent
X
(with the same layer size) encoding stage as additional C`;i,j,k = b` + P`0 ;s(i0 ;i),s(j 0 ;j),s(k0 ;k) w`,`0 ;i0 ,j 0 ,k0 ,
channels to the decoding layer, doubling the number of `0 ,i0 ,j 0 ,k0
channels. We refer this process as concatenation. (1)
The encoding stage consists of a series of ConvNs lay- where P`0 ;s(i0 ;i),s(j 0 ;j),s(k0 ;k) is the input array after the
ers. Let us define the input of a given ConvNs,0 as I`;i,j,k , padding. We sample the convolution sparsely s(i0 ; i) =
where i, j, k ∈ [1, Ns,0 ] is the spatial coordinates, and i × Nst + i0 , and reduce the spatial dimension by a fac-
` ∈ [1, Nch,0 ] is the channel index with Nch,0 being the tor of 23 at each step by choosing the spatial interval
total number of channels. To accommodate the convo- Nst = 2 (strides hereafter). Accompanying the reduc-
lution at the edge, we have added the buffer around the tion of spatial dimension, we increase the number of
input array (padding process). As we use 5 × 5 × 5 con- channels Nch by a factor of 2 at each step of the convo-
volution filter, it suffices to add Np = 2 padding pixels lution, from 128 (Conv64) to 2048 (Conv4). Note that
at both edges of each dimension. We fill the padding the convolution filter w`,`0 ;i0 ,j 0 ,k0 and bias b` are trainable
pixels by reflecting the inner two pixels next to the edge parameters which we adjust for the training.
pixels. The padding and convolution processes are linear op-
After the padding, we apply a three-dimensional con- erations, so any combinations of these operations sim-
volution with a multi-channel filter w`,`0 ;i0 ,j 0 ,k0 and bias plify to a single linear algebra operation. In order to
b` , with indices i0 , j 0 , k 0 ∈ [1, Nk = 5], ` ∈ [1, Nch,1 ], fully utilize the multiple hidden layers of Deep Learn-
ing, we apply the rectified linear unit (ReLU; Hahnloser
et al. 2000; Glorot et al. 2011),

A`;i,j,k = max (C`;i,j,k , 0) , (2)


Local Dark-Matter Map from Deep Learning 5

as a non-linear activation function for each hidden layer. starts from 2 channels of 1283 -grid input layers, and ends
Finally, we apply the batch normalization (Ioffe & with the 2,048 channels of the 43 -grid layer (Conv2). The
Szegedy 2015) final output layers are 643 and 1283 for, respectively,
A`;i,j,k − µ`;i,j,k TNG100 and TNG300. We have also tested other CNN
O`;i,j,k = γ`;i,j,k 2 + β`;i,j,k , (3) architectures with various channel sizes, and confirmed
σ`;i,j,k +
that the CNN architecture that we use here (shown in
to obtain an output ConvNs,1 layer, O`;i,j,k (i, j, k ∈ Figure 1) performs the best among the tested cases.
[1, Ns,1 = Ns,0 /2], ` ∈ [1, Nch,1 ]). Here, µ`;i,j,k and
σ`;i,j,k are the mean and standard deviation of A`;i,j,k
3.2. Training
over samples in a same mini-batch, and  = 10−3 is a
small value for the numerical stability. Note that the We divide the training and validation samples from
mini-batch refers to the bundle of input-output pairs TNG100 so that all sub-cubes from the validation sample
that we have used for updating the trainable param- do not overlap with those from the training sample. As a
eters. The normalization factor γ`;i,j,k and bias factor result, we only use 525 sub-cubes — 432 for training and
β`;i,j,k are another trainable parameters. The batch nor- 93 for validation. For each sub-cube, we make two 643
malization introduces extra level of non-linearity ensur- uniform grids as a two-channel input layer; each channel
ing that the trainable parameters introduced at earlier stores the number of target galaxies (Ngal ) and the av-
hidden layers still affect the output. eraged radial peculiar velocity (Vpec ) in units of km/s.
The decoding stage consists of a series of UpConvNs For the input layer, we apply the same Galactic-latitude
layers, which are constructed in a parallel manner. In mask as the CF3 data (masking out |b| < 10◦ ). For the
contrast to the ConvNs , where we decreases the spatial output layer, we normalize the logarithm of dark-matter
dimension by sparsely sampling the convolved array, we density to be
increase the spatial dimension of each UpConvNs layer:
1
U`;i,j,k = I`;u(i),u(j),u(k) , (4) y= log10 (ρ/ρ0 ) , (5)
4.5
by duplicating the input array I`;i,j,k . Here, u(x) =
dx/Nu e, and we set the upsampling factor Nu = 2 in where ρ0 is the mean dark matter density of the Universe
order to increase the spatial size of U`;i,j,k by a fac- so that all values in the output layer would be between
tor of 8. After the upsampling, we concatenate the −1 and +1.
ConvNs layer (the same size), and apply batch normal- For data augmentation, we allow swapping the
ization. We then apply a three-dimensional convolution (x, y, z)-axes of each sub-cube, which increases the num-
with (Nk , Nst ) = (3, 1), after the reflective padding the ber of samples by a factor of three. We further increase
edge arrays with Np = 1. We decrease the number of the sample size by flipping the axis direction, with which
output channels of each UpConvNs from 1024 to 128 by the number of samples increases eight times. Note that,
a factor of 2. After the convolution, we apply the ReLU unlike U-Net or V-Net, we do not split a single cube
activation function. into multiple smaller cubes for data augmentation be-
In addition to the usual steps described above, the cause that would change of the Galactic-latitude mask
final Output layer requires following two special treat- and the radial peculiar velocity. In the end, we obtain
ments so that the output layer represents the single 10,368 and 2,232 samples, respectively, in training and
dark-matter density proportional to log10 (ρ/ρ0 ) which validation sets.
can be both positive and negative. First, instead of a We implement our CNN architecture in the Keras
gradual decrease of the number of output channels by (Chollet et al. 2015) with the Tensorflow backend
a factor of 2, we set the number of output channels for (Abadi et al. 2015) and perform the training with the
Output as 1. Second, instead of the ReLU activation NVIDIA Tesla V100 graphic processing unit (GPU)
function, whose output range is [0, + inf), we use the with 16GB memory. We choose the mean squared er-
hyperbolic tangent function (tanh) so that its output ror (MSE) as the loss function that the DL minimizes
range becomes finite ([−1, +1] in this case). during the training:
We have adopted different spatial size of the hidden n
layer for TNG100 and TNG300 to accommodating the 1X
LTNG100 = (yi,pred − yi,truth )2 (6)
difference in their spatial resolution. For TNG100, the n i=1
encoding stage starts from 2 channels of 643 -grid in- n  2
1X 1
put layers, and ends with the 2,048 channels of the 23 - = log10 (ρi,pred /ρi,truth ) , (7)
n i=1 4.5
grid layer (Conv2), and, for TNG300, the encoding stage
6 Hong et al.

where the subscripts (i, pred) and (i, truth) are, respec- We perform a similar training for the TNG300 out-
tively, the prediction and truth values of the y (defined come, except for the following differences. First, we have
in Eq. 5) at i-th grid. 10629 training sub-cubes and 1256 validation sub-cubes,
Initially, we set the trainable parameters in the convo- with each sub-cube having 1283 -grids. Unlike TNG100,
lution filters (θ; parameter vector hereafter) randomly. we do not apply further data augmentation, mainly due
The training process for minimizing the loss function is to the expensive computational cost from large CNN
done with 200 epochs, a unit process that updates the architecture size. Second, since the dynamic range of
parameter vector from a subset of the train set and ap- dark-matter density of TNG300 is wider than TNG100,
plies the updated parameter vector to a subset of the we use
1
validation set. The parameter vector update process at y = log10 (ρ/ρ0 ) (11)
each epoch consists of 1728 mini-batches. We set the 5
mini-batch size as 6, mainly due to the GPU memory for the output layer instead. As a result, the MSE loss
limit. For each mini-batch we numerically calculate the function becomes
n  2
gradient of the loss function (∇θ L) and update the pa- 1X 1
LTNG300 = log10 (ρi,pred /ρi,truth ) . (12)
rameter vector by the Adam optimizer (Kingma & Ba n i=1 5
2014),
Third, instead of using a fixed learning rate, we apply
mt /(1 − β1t ) a triangular cyclic learning rate (Smith 2015),
θ t = θ t−1 − αp (8)
v t /(1 − β2t ) +  αU − αL
αt = αL + × min {(t mod T ), T − (t mod T )} ,
mt = β1 mt−1 + (1 − β1 )∇θ Lt (θ t−1 ) (9) T /2
2 (13)
v t = β2 v t−1 + (1 − β2 ) [∇θ Lt (θ t−1 )] . (10)
to avoid the training to be stuck in local minima. Here
Here t is a mini-batch step number starting from zero, T is the number of mini-batches that consists a single
mt and v t are the first and second-moment vectors with learning rate cycle, and we set it as 8. αL and αU are
initial values m0 = v 0 = 0, β1 = 0.9 and β2 = 0.999 the minimum and maximum values of the learning rates,
are exponential decay rates for moment estimates, and respectively. To find a suitable range of learning rates,
 = 10−7 is a small value for the numerical stability. α is we have performed an additional test training with a
the learning rate that determines how fast one updates few epochs by varying learning rates (see Figure 2). If
the parameter vector, and we set it as 10−3 . As a result, the learning rate is too low, i.e., if the parameter vector
the training process for TNG100 takes about 73 hours update is too slow, the loss function as a function of
for a single run. learning rate (L(α)) has a flat slope. On the other hand,
if the learning rate is too high, i.e., if an interval of
1.0
parameter vector update is too large to find a solution,
L(α) presents a noisy increment. We found that 3 ×
0.8
10−8 < α < 4 × 10−5 is a suitable range of the learning
rate and set (αL , αU ) = (3 × 10−8 , 4 × 10−5 ) for the
Loss Function

0.6
triangular cyclic learning rate accordingly. Finally, due
to the large CNN architecture size, we use four NVIDIA
0.4
Tesla V100 GPUs with 32GB memory per each, with
a mini-batch size of 8. For each training, we run 400
epochs by using only 157 mini-batches per epoch, and
0.2
it takes about 90 hours for a single run.
Figure 3 shows the evolution of the MSE loss func-
10−8 10−6 10−4 10−2 100
Learning Rate
tions from both train and validation sets as a function
of epoch in TNG300. Both train and validation losses
Figure 2. Evolution of loss function (L) as a function of similarly decrease over epoch until the validation loss
learning rate of Adam optimizer (α) from an additional test reaches its minimum around 8 × 10−3 at ∼ 140 epochs,
training for TNG300. Too low learning rate (α . 10−8 ) gives while the train loss continues decreasing at all epochs.
a too slow update of the parameter vector, which is pre- Similar minimum values of validation losses have been
sented as a flat slope of L(α). On the other hand, too high found during our test training, and we expect that the
learning rate (α & 10−5 ) prevents finding a solution, which
above value is close to the global minimum of the val-
is presented as a noisy increment of L(α).
idation loss function in our current CNN setup. If the
Local Dark-Matter Map from Deep Learning 7

10−1
use the following four methods for the performance
Train test—visual comparison, joint probability distribution,
Validation histogram, and two-point correlation function (2pCF)
ξ(r) = hδ(x)δ(x + r)ix . To examine the performance of
the each model, we use the Kolmogorov-Smirnov statis-
MSE

tics of the 2pCFs between truth and prediction at a given


scale,
10−2

KS(ξpred , ξtruth ) = max


0
|P̃pred (ξ 0 ) − P̃truth (ξ 0 )| . (14)
ξ

100 101 102


Epoch Here, P̃ (ξ 0 ) = N (ξ < ξ 0 )/Nsample is the empirical distri-
bution function, where Nsample and N (ξ < ξ 0 ) are, re-
Figure 3. Evolution of loss functions from train (blue) and spectively, the number of whole samples and the number
validation (orange) sets as a function of epoch for TNG300. of those satisfying ξ < ξ 0 . The smaller KS(ξpred , ξtruth )
indicates that the predicted probability distribution of
validation loss greatly increases over epoch after reach- the 2pCF is closer to the true distribution, so we use
ing its minimum while the train loss keeps decreasing, it that as a metric to compare the performance of models.
may infer that the learning process starts overfitting the For both TNG100 and TNG300, the models at the mini-
data—the learning process tries to memorize the data mum training loss provide the closest distribution of the
without finding any global feature. In our runs, how- 2pCF predictions to their truth, and we adopt them as
ever, the validation loss does not increase more than 1.1 our optimal models.
times its minimum until the last epoch, which suggests Table 2, Figures 4 and 5 show a visual inspection and
that our result would not suffer from overfitting prob- the statistics of the TNG300 validation samples, which
lem significantly. From each run, we select three models show a good agreement with their true values. Inter-
from three different epochs for the following performance estingly, the predicted dark matter distribution shows
test: at the minimum validation loss, at the minimum small-scale filamentary structures, which are not ap-
training loss, and the last epoch. parently shown in Ngal alone. This is the first indica-
For TNG300, we perform six additional alternative tion of the importance of the (radial) peculiar velocity
training with different configurations of the input layer field for reconstructing the small-scale filamentary struc-
(comparison models hereafter) to understand how such tures; that is, the recovered dark-matter map shows
difference affects our prediction (see Table 1). 16mag much more detailed structure than simply connecting
and 17mag use the alternative absolute B-band magni- the galaxy positions, since the peculiar velocity could
tude cutoffs MB < −16 and −17, respectively. stellar- provide information about the underlying gravitational
Mass uses the logarithm of the total stellar mass rather potential. Simply put, we use the galaxies as test par-
than the simple galaxy number as an input layer, while ticles for recovering the local gravitational field. Note
noVpec does not use the radial peculiar velocity . Fi- that, however, there is a slight difference in the detailed
nally, DMhalo uses the dark matter halos in the dark- distribution of filamentary structures between truth and
matter-only simulation TNG300-1-Dark instead of galax- prediction. Also, note that there exists a sharp lower cut
ies in the TNG300-1.1 in the predicted density min ρpred ∼ 10−2 ρ0 . The above
two issues could be overcome by using higher-resolution
4. RESULTS
hydrodynamic simulations and observational data with
4.1. Performance Test more low-brightness galaxies. Also, fine-tuned choices of
To test the model parameters tuned with TNG100 and loss function might help manage an issue about a slight
TNG300 training sets, we apply the model to the val- difference of filamentary structures.
idation samples to compare the resulting dark-matter After choosing the optimal models, we perform the
density cube with the ground truth. Specifically, we convergence test between models with different simula-
tion resolutions and setups. First, we compare the local
dark-matter density field predictions from TNG100 and
1 Note that Modi et al. (2018) performed a similar study to re- TNG300 within the radius r = 10 Mpc/h. We find that
construct the (initial) density perturbation from the dark matter
halo distributions by DL, while they focused more on large scales
they show similar distribution up to r ∼ 4 Mpc/h, while
such as baryon acoustic oscillation (BAO) rather than relatively the dark-matter map from TNG100 shows finer small-
small scales such as ours. scale structures than TNG300 (see Section 4.3). Also,
8 Hong et al.

Table 1. Summary of TNG300 and its comparison models used in this paper. Each comparison model is the same as TNG300
except those mentioned in its “Description.”

Model Description
TNG300 Simulation: TNG300-1 hydrodynamic simulation.
Center galaxies: 4 × 1010 M < M? < 1011 M after resolution correction.
Target galaxies: MB < −15 after resolution correction.
Input layer: 2-channel (Ngal and Vpec ).
Hubble parameter: 67.77km/s/Mpc
16mag Target galaxies: MB < −16 after resolution correction.
17mag Target galaxies: MB < −17 after resolution correction.
noVpec Input layer: 1-channel (Ngal ).
stellarMass Input layer: 2-channel (log10 (M? /M ) (logarithm of the total stellar mass) and Vpec ).
DMhalo Simulation: TNG300-1-Dark dark-matter-only simulation.
Center & target galaxies: applying halo mass cut that matches the same galaxy number density to TNG300.
diffH0 Hubble parameter: 75km/s/Mpc

Table 2. Summary of the performance test done by validation samples of TNG100,


TNG300, and their comparison models. KS(ξpred , ξtruth ) is the Kolmogorov-Smirnov
statistics of the two-point correlation functions of dark-matter distribution between truth
and prediction. EAGLE-TNG100 is the application of the TNG100 model to the EAGLE
samples. diffH0 is identical to TNG300 since Hubble flow estimation is not considered in
this test.

KS(ξpred , ξtruth )
Model log10 (ρpred /ρtruth )
0 − 1 Mpc/h 1 − 3 Mpc/h 3 − 10 Mpc/h
TNG100 −0.014 ± 0.543 0.263 ± 0.035 0.175 ± 0.087 0.130 ± 0.042
EAGLE-TNG100 +0.129 ± 0.491 0.171 ± 0.055 0.152 ± 0.047 0.149 ± 0.040
TNG300 −0.020 ± 0.451 0.153 ± 0.035 0.134 ± 0.040 0.163 ± 0.017
16mag −0.008 ± 0.468 0.109 ± 0.010 0.161 ± 0.033 0.254 ± 0.016
17mag +0.017 ± 0.481 0.143 ± 0.037 0.168 ± 0.018 0.251 ± 0.019
noVpec +0.016 ± 0.481 0.367 ± 0.115 0.407 ± 0.061 0.170 ± 0.036
stellarMass −0.050 ± 0.471 0.186 ± 0.056 0.218 ± 0.016 0.269 ± 0.021
DMhalo +0.002 ± 0.481 0.264 ± 0.029 0.243 ± 0.030 0.263 ± 0.034

we apply the CNN model from TNG100 to the test sam- tion (DMhalo) has slightly more offset in the distribution
ple of EAGLE (EAGLE-TNG100 in Table 2). of 2pCFs. Those without using the radial peculiar veloc-
Note that we do not apply the CNN model from ity as inputs (noVpec), however, do not reproduce any
TNG300 to EAGLE because the volume of EAGLE is not small-scale filamentary structure shown in the true dark
sufficiently larger than the volume of TNG300 sub-cubes. matter distribution (see the right panel of Figure 4).
We find that its performance test result is similar to the From its visual inspection, one could interpret the out-
TNG100 validation sample, except that EAGLE-TNG100 put of noVpec as a smoothing of the galaxy number
tends to slightly overestimate the dark-matter density distribution—the only available input of the given DL
(see Figure 5 and Table 2). model— with a few Mpc-scale. As a result, the 2pCFs of
We also test the performance of various comparison noVpec show a significant deviation from their truth in
models of TNG300 (see Table 1 for definitions). Most small scales with r . 3 Mpc/h (see Table 2). From the
comparison models show similar overall performance to comparison to TNG300 and its other comparison models,
TNG300, while those from the dark-matter-only simula- it is apparent that the (radial) peculiar velocity plays a
Local Dark-Matter Map from Deep Learning 9

Ngal Vpec Truth TNG300 noVpec

10 0.4

Ngal
0
0.2
−10
YZ plane
0.0

Vpec [km/s]
10
[h−1 Mpc]

0 0
−10
XZ plane
−100
10
101

ρ/ρ0
0
−10
YX plane
10−1
−10 0 10 −10 0 10 −10 0 10 −10 0 10 −10 0 10
[h−1 Mpc]

Figure 4. 3-way projections of a single TNG300 validation sample with 5 Mpc/h-thickness. From left to right: galaxy number
(Ngal ), radial peculiar velocity (Vpec ), truth dark-matter density (ρtruth ), reconstructed dark-matter density (ρTNG300 ), and
another reconstruction from the CNN architecture without using the radial peculiar velocity (noVpec; ρnoVpec ). TNG300 can
well reconstruct the filamentary structure of a few-Mpc scales in the true dark-matter distribution, while noVpec does not show
such structure.
significant role in reconstructing the small-scale filamen- Local Void (Tully & Fisher 1987) is also apparent (also
tary structure. shown on the SGZ-SGX plane), which might extend be-
yond the boundary of our local universe sample. In Fig-
4.2. Three-dimensional view of the Local Cosmic Web
ure 6, we also present the velocity flow lines derived
Figure 6 shows a sliced view of the reconstructed Cos- from the reconstructed gravitational potential gradient
mic Web integrated over 4 Mpc/h thickness. Each panel with arrows and black lines. The velocity flow shows
shows the Cosmic Web on the plane of the Supergalactic the motion of material from the Local Void to nearby
Cartesian coordinates (SGX, SGY, and SGZ), extended filamentary structures and clusters such as the Local
to the full cube with the side length of 40 Mpc/h. Fig- Sheet, Fornax Wall, and Virgo cluster. Note that we
ure 6 clearly shows known local objects that we des- cannot reproduce the velocity flow from the Virgo clus-
ignated by their common name. The figure also re- ter to the Great Attractor (+SGX-direction), because
covers known local large-scale structures. For example, of the limited extension of the volume that we analyze
we find a 10 Mpc/h-spread along +SGY-direction in the here. However, we would like to emphasize that the re-
SGZ-SGY (upper left panel) and SGY-SGX (lower-right covered dark-matter map provides us detailed density
panel) planes. This structure is known as Local Sheet, and velocity fields around these known local large-scale
which connects the Local Group and Virgo cluster and structures.
contains M81, NGC5194, Canes II, and Coma I groups The recovered Cosmic Web also shows a hint of new
(Tully et al. 2008; Courtois et al. 2013). We also find structures that require further investigation. For exam-
that, around the Local Group, the Local Sheet is con- ple, the direction of the Local Sheet is similar to the
nected to the Fornax Wall (Fairall et al. 1994), which is a direction of the so-called vast polar structure (VPOS),
20 Mpc/h-sized spread along (−SGY,−SGZ)-direction, which consists of satellite galaxies, globular clusters and
containing Fornax cluster, Eridanus cluster, and Dorado stellar streams around the Milky-way galaxy (Pawlowski
group as members (upper-left panel). At the opposite et al. 2012). As shown in Figure 6, the Local Sheet, be-
direction to the Fornax Wall on the SGZ-SGY plane, the
10 Hong et al.

103
105
0
10
TNG300 Truth
103 −1 102 Prediction
10

hδ(x)δ(x + r)i
df /d log10 ρ
101
ρpred /ρ0

10−2 101

10−1 10−3
100
−3
10 10−4

10−5 −5 10−5 −5 0
10 10−2 101 104 10 10−2 101 104 100 101
ρtruth /ρ0 ρ/ρ0 −1
r [h Mpc]

105 103
100
EAGLE Truth
103 10 2
10−1 Prediction

hδ(x)δ(x + r)i
df /d log10 ρ

101
ρpred /ρ0

10−2
101

10−1 10−3
100
10−3 10−4

10−5 −5 10−5 −5 0
10 10−2 101 104 10 10−2 101 104 100 101
ρtruth /ρ0 ρ/ρ0 r [h−1 Mpc]

Figure 5. Result of the performance tests for the deep learning result using the three-dimensional dark-matter density field
of simulations. Top panel: statistical comparison between the ground truth and the predicted dark-matter density from the
entire TNG300 validation sample. From left to right: joint probability distribution (colors) with 1, 2, 3-σ certainty level contours
(lines), median (lines) and 1-σ deviation (shades) of histograms, and median (lines) and 1-σ deviation (shades) of the two-point
correlation functions. Bottom panel: similar to the top panel, but by applying the TNG100 training to the entire EAGLE test
sample.

ing the strongest filamentary structure around the Local various samples (Tully et al. 2013), Tully-Fisher relation
Group, is a source of velocity flow; that might cause a using Spitzer [3.6] photometry, and the Fundamental
connection between the two. Also, a couple of small Plane relation from the Six Degree Field Galaxy Survey
filaments are visible in our maps, which could be good (6dFGS ) (Tully et al. 2016). We then generate 1,000
targets for systematic examination with deep imaging sets of random distance moduli that follow the normal
surveys. distribution,
Furthermore, to estimate the uncertainties of the
∆µ2
 
dark-matter map, we perform a stress test on our CNN 1
P (∆µ) = √ exp − 2 . (16)
models by incorporating distance measurement uncer- µ 2π 2µ
tainties in the CF3. We use the one standard deviation
uncertainty in distance modulus (µ ) in the CF3, Then, we re-calculate the radial peculiar velocity by
s subtracting the Hubble flow corresponding to the ran-
1 dom distances from the VGSR . Since the distance mea-
µ ≡ P 2 . (15) surement error exists only along the radial direction,
i 1/i
we have generated the two-dimensional column density
Here i includes the one standard deviation uncertainty map of the dark matter that is less affected by the er-
determined from a recalibration of galaxy magnitude ror than the three-dimensional dark-matter density field
with H I linewidth (Tully & Courtois 2012), distance (see Figure 9). Also, we find that the dark-matter col-
measurement of the Tip of the Red Giant Branch from umn density map driven from TNG300 shows signifi-
the Hubble Space Telescope, Type Ia supernovae from cantly less deviation than that of TNG100, which suf-
Local Dark-Matter Map from Deep Learning 11
20

15
NGC3607

10 Coma I

Leo I
5 NGC5194
SGY [Mpc/h]

M81
0 Local Void

−5

−10 Fornax Cluster

−15

−20
−20 −15 −10 −5 0 5 10 15 20
SGZ [Mpc/h]

20 20

15 15

10 10

Ursa Major N
5 5
Ursa Major S
M81 M81 Canes II
SGX [Mpc/h]

SGX [Mpc/h]

0 Local Void 0

NGC5128 NGC5128 Virgo Cluster


−5 −5
NGC7582
NGC4697

−10 −10

−15 −15

−20 −20
−20 −15 −10 −5 0 5 10 15 20 −20 −15 −10 −5 0 5 10 15 20
SGZ [Mpc/h] SGY [Mpc/h]

Figure 6. Three-dimensional density maps of the local dark matter with 40 Mpc/h-boxsize and 4 Mpc/h-thickness. ‘X’-mark
at the center: Milky-way galaxy. Dots: galaxies with MB < −15. Texts: galaxy groups, clusters, and local structures. Arrows:
estimated directions of motion derived from the gradient of the reconstructed gravitational potential.

fers from some spurious structure consistently appearing where θ, r, ρ(θ, r) are the two-dimensional sky coordi-
near the Galactic plane. nates, distance from the observer, and the dark-matter
density at the given (θ, r), respectively. We use the
4.3. Sky map of the Local Cosmic Web Healpix (Zonca et al. 2019; Górski et al. 2005) pack-
The left panels of Figure 7 (labeled as TNG300) show age to reconstruct the two-dimensional sky map from
the recovered local dark-matter map on the sky (gray the three-dimensional data cube. We set the resolution
map), parameter Nside = 128, which roughly corresponds to
Z the angular resolution of 270 . This figure also shows the
Σ(θ) ≡ dr ρ(θ, r) , (17) locations and radial peculiar velocities of galaxies that
12 Hong et al.

0.7 < r/h−1 Mpc < 4.0 0.7 < r/h−1 Mpc < 4.0
TNG300 TNG100
Canes I
M81
NGC5128
Maffei

Sculptor

4.0 < r/h−1 Mpc < 8.0 0.7 < r/h−1 Mpc < 4.0
TNG300 DMhalo
Canes II Leo I

NGC1023

8.0 < r/h−1 Mpc < 12.0 0.7 < r/h−1 Mpc < 4.0
TNG300 noVpec
Ursa Major S Virgo Cluster

NGC2997
NGC6744
NGC628

12.0 < r/h−1 Mpc < 16.0 12.0 < r/h−1 Mpc < 16.0
TNG300 noVpec
Ursa Major N NGC3607
NGC4038

Fornax Cluster
NGC7582

16.0 < r/h−1 Mpc < 20.0 16.0 < r/h−1 Mpc < 20.0
TNG300 noVpec
NGC5485 NGC5746

NGC3923

Eridanus Cluster
IC1459

−1.0 −0.5 0.0 0.5 1.0 −100 −50 0 50 100


log10 Σ/Σ0 Vpec [km/s]

Figure 7. Two-dimensional full-sky map of the local dark-matter column density with 4 Mpc/h widths. Left panels: predictions
from TNG300 training, from the nearest to the farthest radial bin. Right panels: comparison predictions from TNG100 training
(TNG100), training with dark matter halos from the dark matter-only simulation (DMhalo), and training without using the
radial peculiar velocity (noVpec). Small dots: positions and peculiar velocity (color) of known local galaxies. Large dots: galaxy
groups and clusters with their names.
Local Dark-Matter Map from Deep Learning 13

we use for the reconstruction (color-coded dots), as well radial bin well represent the spatial resolution of the
as the locations of some well-known galaxy groups and three-dimensional grid (0.3215 Mpc/h). On the other
clusters (large dots). hand, δ` at the farthest radial bin may mean a typi-
cal scale of the filamentary structure at given radial bin
width and galaxy number density. For the statistical
0.04
analysis, we degrade the angular resolution of each map
Angular Covariance Function

to δθ0 —Nside = 4, 8, 8, 16, and 16 from the nearest to


0.03 the farthest radial bins— and assume that each pixel in
the degraded map is statistically independent.
0.02 Figure 9 shows the mean (hlog10 Σi; left panel) and
0.7 < r/h−1 Mpc < 4 the standard deviation (∆ log10 Σ; middle panel) of the
4 < r/h−1 Mpc < 8
0.01 logarithm of the local dark-matter column density over
8 < r/h−1 Mpc < 12
12 < r/h−1 Mpc < 16
1,000 realizations incorporating the uncertainties of the
16 < r/h−1 Mpc < 20 distance measurement. We find that the standard devi-
0.00
100 101
ation per pixel stays in the range of ∆ log10 Σ/Σ0 '
Angular Distance [deg] 0.1 − 0.4, with only a mild dependence to the den-
sity contrast. As a result, the signal-to-noise ratio
Figure 8. Angular covariance function C(δθ) as a function SNR ≡ |hlog10 Σi|/∆ log10 Σ scales almost linearly as
of angular distance (δθ) in the sky. Each angular covariance the density contrast, reaching up to SNR '10 for the
function roughly follows C(δθ) ∝ exp(−δθ/δθ0 ), where δθ0 density peaks. On average, the signal-to-noise ratios
is a proxy of angular resolution of a degraded map each of for dark-matter distribution per pixel at higher Galac-
whose pixel is statistically independent.
tic latitudes (|b| > 10◦ ) are 4.25, 3.76, 3.94, 4.19, and
4.52, respectively, from the nearest to the farthest radial
The map in Figure 7 uses the radial distance and bin.
radial peculiar velocities reported in the Cosmicflows- Note that, in addition to the distance measurement
3 catalog (Tully et al. 2016). We have mitigated the uncertainty, there are systematic uncertainties in DL
10 − 30% uncertainties of distance measurement in the mapping itself into the error budget. For example, the
catalog by adopting the radial binning of ∆r = 4 Mpc/h. galaxy simulations with different resolutions or different
We further analyze the statistical uncertainties of the sub-grid prescriptions can lead to different DL mapping.
recovered dark-matter map by generating 1,000 real- We check such systematic effect by comparing TNG300
izations incorporating the uncertainties of the distance with various comparison models, including those already
measurement (see Section 4.2). From the high angular introduced in Table 1. To do this, we calculate the on-
resolution map (Nside = 128), we define the angular sky average of the systematics
covariance function,
| log10 Σ − log10 ΣTNG300 |
hδΣ(θ) δΣ(θ 0 )iN,θ,θ0 ∆sys ≡ , (20)
C(δθ) ≡ , (18) ∆ log10 ΣTNG300
Σ20
where Σ and ΣTNG300 are the local dark-matter column
where Σ0 = ρ0 ∆r is the mean dark-matter column den- densities from a given comparison model and TNG300,
sity, h. . .iN,θ,θ0 is the average over N =1,000 realizations both by adopting the reported values of galaxy loca-
and sky coordinates θ and θ 0 that satisfy |θ − θ 0 | = δθ, tions and peculiar velocities. First, we check the sys-
and δΣ(θ) ≡ Σ(θ) − hΣ(θiN . We found that the an- tematic effect of the resolution by comparing the local
gular covariance function follows an exponential decay dark-matter map estimated from TNG100 and TNG300.
over δθ,   The top-right panel of Figure 7 shows the r < 4 Mpc/h
δθ
C(δθ) ≈ C0 exp − , (19) bin dark-matter map driven from the high-resolution re-
δθ0 sult (TNG100). TNG100 systematically underestimate
and the values of the angular scale that shows a strong the density contrast by ∆sys = 2.3 on average (see Ta-
pixel-to-pixel correlation are δθ0 = 20.7◦ , 9.71◦ , 6.53◦ , ble 3).
5.04◦ , and 4.24◦ , respectively, from the nearest (r < Secondly, to estimate the systematic effect from differ-
4 Mpc/h) to the farthest (16 Mpc/h < r < 20 Mpc/h) ent sub-grid prescriptions, we have repeated the deep-
radial bins (see Figure 8). δθ0 at different radial bins learning procedure by using the dark-matter halo sam-
correspond to the linear scales δ` = 0.26, 0.68, 0.92, ples from the dark-matter-only simulation TNG300-
1.06, and 1.18 Mpc/h, respectively. δ` at the nearest 1-Dark by matching the galaxy/halo number density
14 Hong et al.

Table 3. On-sky average (median and 1-σ certainty level in the parenthesis) of the systematics ∆sys ≡ | log10 Σ −
log10 ΣTNG300,face |/∆ log10 ΣTNG300 over high Galactic latitude |b| > 10◦ with different radial bins. See Table 1 for the defini-
tion of each comparison model except TNG100.

Comparison Model 0.7 − 4 Mpc/h 4 − 8 Mpc/h 8 − 12 Mpc/h 12 − 16 Mpc/h 16 − 20 Mpc/h


TNG100 2.281 (1.837+1.993
−1.104 ) 1.474 (1.196+1.414
−0.842 ) - - -
diffH0 0.212 (0.171+0.223
−0.115 ) 0.162 (0.133+0.148
−0.092 ) 0.154 (0.116+0.161
−0.083 ) 0.152 (0.117+0.153
−0.082 ) 0.160 (0.128+0.151
−0.092 )
+0.748 +1.089 +0.729 +0.751
16mag 1.032 (0.949−0.647 ) 1.093 (0.868−0.611 ) 0.862 (0.716−0.508 ) 0.785 (0.641−0.455 ) 0.804 (0.631+0.790
−0.443 )
17mag 1.178 (0.901+1.081
−0.572 ) 1.105 (0.889+1.026
−0.621 ) 1.001 (0.815+0.947
−0.575 ) 0.887 (0.726+0.862
−0.502 ) 0.898 (0.734+0.833
−0.506 )
noVpec 1.935 (1.715+1.919
−1.359 ) 1.105 (0.834+1.120
−0.631 ) 0.943 (0.701+0.890
−0.524 ) 0.828 (0.672+0.751
−0.470 ) 0.750 (0.626+0.742
−0.440 )
stellarMass 1.544 (1.256+1.435
−0.843 ) 1.175 (0.946+1.156
−0.684 ) 0.925 (0.734+0.909
−0.521 ) 0.877 (0.692+0.837
−0.485 ) 0.907 (0.713+0.899
−0.490 )
DMhalo 1.737 (1.154+2.253
−0.863 ) 1.445 (1.127+1.414
−0.816 ) 1.176 (0.913+1.097
−0.610 ) 1.057 (0.846+1.029
−0.595 ) 0.957 (0.796+0.889
−0.574 )

(DMhalo). The right panels of Figure 9 show the dif- radial bin. We, however, anticipate that the theoret-
ference between the two dark-matter maps in units of ical uncertainties for the DL mapping would be most
standard deviation at each pixel. Even with this extreme substantial for this region. For example, from the afore-
comparison between full hydrodynamic simulation and mentioned studies on systematic uncertainties, we find
pure N -body simulation , we find that systematic effects that, on average, lower Galactic latitudes (|b| < 10◦ )
lead to ∆sys = 1.7, 1.4, 1.2, 1.1, and 1.0 on average from map suffers about δ∆sys ' 0.5 more systematical shifts
the top (nearest) to the bottom (furthest) maps. than higher Galactic latitudes (|b| > 10◦ ) map. This
We further test the systematic effect due to different is indicated in the top two panels of Figure 7 and the
Hubble parameters (H0 = 75 km/s/Mpc; diffH0) and systematic shifts shown in the right panels of Figure 9.
find only ∆sys ' 0.15. Different B-band magnitude cuts
5. DISCUSSION
(MB < −16 and −17; 16mag and 17mag, respectively)
and using total stellar mass instead of galaxy number In this paper, we present a novel convolutional neural
(stellarMass) lead to ∆sys ' 1. Most importantly, none network (CNN) -based deep learning (DL) method of re-
of the systematic maps shows a significant correlation constructing the local dark-matter distribution map and
with the derived cosmic web structure, ensuring the ro- discover the local Cosmic-Web structure traced by the
bustness of the derived dark-matter distribution, or the positions and radial peculiar velocities of Cosmicflow-3
Cosmic Web (see the right panel of Figure 9). galaxies. We find that including the radial peculiar ve-
The most striking feature that we have recovered in locity field is the key to recover the dark matter distribu-
this study is the filamentary Cosmic Web that is appar- tion in the Cosmic Web. Incorporating the observational
ent in Figures 7 and 9. First of all, we find that the radial uncertainties in the galaxy distance measurements, the
peculiar velocity information is vital to reconstructing average detection significance of the dark-matter map
the cosmic web, without which the same DL algorithm exceeds 4.1-σ for each Healpix pixel at higher Galac-
can not reproduce the Cosmic Web structure at all. tic latitudes (|b| > 10◦ ). The quoted statistical signifi-
For example, the right panels in Figure 7, indicated by cance, however, does not include the uncertainties in the
noVpec, show the deep-learning result only using galaxy galaxy-to-dark-matter mapping itself. We have tested
distributions. Note the absence of the filamentary struc- that the DL results stay robustly for three different sim-
ture in those maps. We note that the noVpec maps re- ulations: TNG100-1 and TNG300-1 from the Illustris-
semble the smoothed version of the galaxy distribution. TNG simulation and RefL0100N1504 from the EAGLE
The deep-learning algorithm with stellar-mass weighted simulation, but future studies must quantify the theo-
galaxy distribution, without peculiar velocity informa- retical uncertainties by applying the same method to the
tion, leads to the similarly poor quality map. large-scale structure simulations with different baryonic
Another interesting feature in the map is the dark- prescriptions. The comparison of the DL results between
matter distribution at lower Galactic latitudes (|b| < TNG300-1 and N -body simulations, however, indicates
10◦ ) where we do not have any input galaxy data. To that the filamentary Cosmic-Web structure may not suf-
our surprise, we find that the averaged signal-to-noise fer from the systematic effects.
ratios per pixel for this region are 4.18, 4.73, 5.31, 5.80, The main statistical uncertainty in the galaxy data
and 6.21, respectively, from the nearest to the farthest comes from the uncertainty in the distance measure-
ment. As the observed shift in the galaxy spectra con-
Local Dark-Matter Map from Deep Learning 15

Mean Standard Deviation DM-only Systematics


0.7 < r/h−1 Mpc < 4.0 0.7 < r/h−1 Mpc < 4.0 0.7 < r/h−1 Mpc < 4.0

4.0 < r/h−1 Mpc < 8.0 4.0 < r/h−1 Mpc < 8.0 4.0 < r/h−1 Mpc < 8.0

8.0 < r/h−1 Mpc < 12.0 8.0 < r/h−1 Mpc < 12.0 8.0 < r/h−1 Mpc < 12.0

12.0 < r/h−1 Mpc < 16.0 12.0 < r/h−1 Mpc < 16.0 12.0 < r/h−1 Mpc < 16.0

16.0 < r/h−1 Mpc < 20.0 16.0 < r/h−1 Mpc < 20.0 16.0 < r/h−1 Mpc < 20.0

−1.0 −0.5 0.0 0.5 1.0 0.0 0.1 0.2 0.3 0.4 0.5 −3 −2 −1 0 1 2 3
hlog10 Σ/Σ0 i ∆ log10 Σ/Σ0 (log10 ΣDM − log10 Σ)/∆ log10 Σ

Figure 9. Same as Figure 7, but showing statistical maps. Left panels: mean of the logarithm of dark-matter column-density
estimated from 1000 random realizations incorporating the uncertainties in distance estimate to the local galaxies. Middle panels:
standard deviation from 1,000 random realizations (Nside = 4, 8, 8, 16, 16 from top to bottom). Right panels: systematic bias
from different simulation input for the deep-learning (TNG300 vs. DMhalo).
16 Hong et al.

strains the sum of the distance (Hubble flow) and the additional study that applies the existing methods to
radial peculiar velocity, the uncertainty affects both the similar observational and simulation data to ours and
galaxy distribution and the radial peculiar velocity field. compares them to our DL method would be beneficial,
Therefore, to obtain a dark-matter map with higher sig- and we leave it for the future.
nificance, it is necessary to explore the ways to reduce ACKNOWLEDGMENTS
the uncertainties of the current distance estimators such
as the Tip of the Red Giant Branch, the Type Ia su- Authors acknowledge Christophe Pichon, Changbom
pernova, and the Fundamental Plane through continu- Park, Sungryong Hong, Inkyu Park, Dongsu Bak,
ous cross-calibration (Tully et al. 2016), and to increase Graziano Rossi, and Yung-Kyun Noh for discussion. Au-
the number of galaxies with measured distances through thors also acknowledge an anonymous referee for sug-
systematic surveys (e.g., 6dFGS (Springob et al. 2014), gestions to improve this article. The list of nearby
James Webb Space Telescope (Gardner et al. 2006)). galaxy groups and clusters is derived from www.
We anticipate that the reconstructed three-dimensio- atlasoftheuniverse.com. Authors acknowledge the Ko-
nal dark-matter map and peculiar velocity field will open rea Institute for Advanced Study for providing comput-
an entirely new chapter of cosmological study. For ex- ing resources (KIAS Center for Advanced Computation
ample, the dark-matter map can make it possible to Linux Cluster System). Computational data were trans-
run the cosmological galaxy simulations with the pre- ferred through a high-speed network provided by the
cise initial condition of the Local Group for studying Korea Research Environment Open NETwork (KRE-
the past and future of our cosmic neighborhood. It will ONET).
also allow the in-depth study of the nature of dark mat- SEH was partly supported by Basic Science Research
ter by cross-correlating the reconstructed dark matter Program through the National Research Foundation of
map with the full-sky diffuse emission maps constructed Korea funded by the Ministry of Education (2018R1A6-
from the radio-to-gamma-ray electromagnetic spectra as A1A06024977). SEH was also partly supported by the
well as the full-sky map of gravitational wave binaries. project ᄋ
ᅮᄌ ᅮ거ᄃ ᅢ구조ᄅᆯᄋ
ᅳ ᅵᄋᆼᄒ
ᅭ ᆫᄋ
ᅡ ᆷᄒ
ᅡ ᆨᄋ
ᅳ ᅮ주ᄋ ᆫᄀ
ᅧ ᅮ (“Un-
The latter can test the models where black holes in bi- derstanding Dark Universe Using Large Scale Structure
naries have formed out of dark matter (Shandera et al. of the Universe”), funded by the Ministry of Science.
2018). DJ was supported at Pennsylvania State University by
Finally, as we have introduced a novel CNN-based DL NSF grant (AST-1517363) and NASA ATP program
method to reconstruct the local Cosmic Web, the quan- (80NSSC18K1103). JK was supported by a KIAS Indi-
titative study comparing the prediction power of the DL vidual Grant (KG039603) via the Center for Advanced
method presented here with pre-existing methods such Computation at Korea Institute for Advanced Study.
as BORG may be in order. Note that, however, many
previous studies reconstruct the dark matter distribu-
Software: HEALPix (Górski et al. 2005), Healpy
(Zonca et al. 2019), astropy (Astropy Collaboration et al.
tion on sales much larger than the size of our local Cos-
2013, 2018), NumPy (Van Der Walt et al. 2011; Harris
mic Web (e.g., & 3 − 5 Mpc/h in Jasche & Wandelt
et al. 2020), Scipy (Jones et al. 2001; Virtanen et al.
2013; Jasche et al. 2015), which complicates the direct
2020), matplotlib (Hunter 2007), pandas (Wes McKin-
comparison between the two methods. Nevertheless, an
ney 2010), Keras (Chollet et al. 2015), Tensorflow back-
end (Abadi et al. 2015)

REFERENCES
Aaronson, M. 1983, The Astrophysical Journal Letter, 266, Abbott, T. M. C., Abdalla, F. B., Alarcon, A., et al. 2018,
L11, doi: 10.1086/183969 Physical Review D, 98, 043526,
doi: 10.1103/PhysRevD.98.043526
Aartsen, M. G., Ackermann, M., Adams, J., et al. 2018,
Ackermann, M., Albert, A., Anderson, B., et al. 2015,
Eur. Phys. J. C, 78, 831,
Physical Review Letter, 115, 231301,
doi: 10.1140/epjc/s10052-018-6273-3
doi: 10.1103/PhysRevLett.115.231301
Abadi, M., Agarwal, A., Barham, P., et al. 2015, Akerib, D. S., Alsum, S., Araújo, H. M., et al. 2017,
TensorFlow: Large-Scale Machine Learning on Physical Review Letter, 118, 021303,
Heterogeneous Systems. https://www.tensorflow.org/ doi: 10.1103/PhysRevLett.118.021303
Local Dark-Matter Map from Deep Learning 17

Ammazzalorso, S., Gruen, D., Regis, M., et al. 2020, Fairall, A. P., Paverd, W. R., & Ashley, R. P. 1994, in
Physical Review Letter, 124, 101102, Astronomical Society of the Pacific Conference Series,
doi: 10.1103/PhysRevLett.124.101102 Vol. 67, Unveiling Large-Scale Structures Behind the
Anderson, L., Aubourg, É., Bailey, S., et al. 2014, Monthly Milky Way, ed. C. Balkowski & R. C. Kraan-Korteweg,
Notices of the Royal Astronomical Society, 441, 24, 21
doi: 10.1093/mnras/stu523 Fang, K., Banerjee, A., Charles, E., & Omori, Y. 2020,
Arcadi, G., Dutra, M., Ghosh, P., et al. 2018, Eur. Phys. J. arXiv e-prints, arXiv:2002.06234.
C, 78, 203, doi: 10.1140/epjc/s10052-018-5662-y https://arxiv.org/abs/2002.06234
Fornasa, M., Cuoco, A., Zavala, J., et al. 2016, Physical
Astropy Collaboration, Robitaille, T. P., Tollerud, E. J.,
Review D, 94, 123005, doi: 10.1103/PhysRevD.94.123005
et al. 2013, Astronomy & Astrophysics, 558, A33,
Gardner, J. P., Mather, J. C., Clampin, M., et al. 2006,
doi: 10.1051/0004-6361/201322068
Space Science Reviews, 123, 485,
Astropy Collaboration, Price-Whelan, A. M., SipHocz,
doi: 10.1007/s11214-006-8315-7
B. M., et al. 2018, aj, 156, 123,
Giesen, G., Boudaud, M., Génolini, Y., et al. 2015, Journal
doi: 10.3847/1538-3881/aabc4f
of Cosmology and Astroparticle Physics, 2015, 023,
Atlas Collaboration. 2019, Journal of High Energy Physics,
doi: 10.1088/1475-7516/2015/09/023
2019, 142, doi: 10.1007/JHEP05(2019)142
Glorot, X., Bordes, A., & Bengio, Y. 2011, in Proceedings
Bell, E. F., McIntosh, D. H., Katz, N., & Weinberg, M. D. of the fourteenth international conference on artificial
2003, The Astrophysical Journal Supplement Series, 149, intelligence and statistics, 315–323
289, doi: 10.1086/378847 Górski, K. M., Hivon, E., Banday, A. J., et al. 2005, The
Camps, P., Trčka, A., Trayford, J., et al. 2018, The Astrophysical Journal, 622, 759, doi: 10.1086/427976
Astrophysical Journal Supplement Series, 234, 20, Gottloeber, S., Hoffman, Y., & Yepes, G. 2010, arXiv
doi: 10.3847/1538-4365/aaa24c e-prints, arXiv:1005.2687.
Carlesi, E., Sorce, J. G., Hoffman, Y., et al. 2016, Monthly https://arxiv.org/abs/1005.2687
Notices of the Royal Astronomical Society, 458, 900, Hahnloser, R. H. R., Sarpeshkar, R., Mahowald, M. A.,
doi: 10.1093/mnras/stw357 Douglas, R. J., & Seung, H. S. 2000, Nature, 405, 947,
Carrick, J., Turnbull, S. J., Lavaux, G., & Hudson, M. J. doi: 10.1038/35016072
2015, Monthly Notices of the Royal Astronomical Harris, C. R., Millman, K. J., van der Walt, S. J., et al.
Society, 450, 317, doi: 10.1093/mnras/stv547 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2
Chollet, F., et al. 2015, Keras, https://keras.io Huchra, J. P., Macri, L. M., Masters, K. L., et al. 2012, The
Astrophysical Journal Supplement Series, 199, 26,
Clowe, D., Bradač, M., Gonzalez, A. H., et al. 2006, The
doi: 10.1088/0067-0049/199/2/26
Astrophysical Journal Letter, 648, L109,
Hunter, J. D. 2007, Computing in Science Engineering, 9,
doi: 10.1086/508162
90, doi: 10.1109/MCSE.2007.55
Cooke, R. J., Pettini, M., Jorgenson, R. A., Murphy, M. T.,
Ioffe, S., & Szegedy, C. 2015, arXiv e-prints,
& Steidel, C. C. 2014, The Astrophysical Journal, 781,
arXiv:1502.03167. https://arxiv.org/abs/1502.03167
31, doi: 10.1088/0004-637x/781/1/31
Jasche, J., Leclercq, F., & Wandelt, B. D. 2015, Journal of
Courtois, H. M., Pomarède, D., Tully, R. B., Hoffman, Y.,
Cosmology and Astroparticle Physics, 2015, 036,
& Courtois, D. 2013, The Astronomical Journal, 146, 69,
doi: 10.1088/1475-7516/2015/01/036
doi: 10.1088/0004-6256/146/3/69
Jasche, J., & Wandelt, B. D. 2013, Monthly Notices of the
Crain, R. A., Schaye, J., Bower, R. G., et al. 2015, Monthly Royal Astronomical Society, 432, 894,
Notices of the Royal Astronomical Society, 450, 1937, doi: 10.1093/mnras/stt449
doi: 10.1093/mnras/stv725 Jeffrey, N., Lanusse, F., Lahav, O., & Starck, J.-L. 2020,
Davis, M., Efstathiou, G., Frenk, C. S., & White, S. D. M. Monthly Notices of the Royal Astronomical Society, 492,
1985, The Astrophysical Journal, 292, 371, 5023, doi: 10.1093/mnras/staa127
doi: 10.1086/163168 Jones, E., Oliphant, T., Peterson, P., et al. 2001
Desjacques, V., Jeong, D., & Schmidt, F. 2018, Physics Kingma, D. P., & Ba, J. 2014, arXiv e-prints,
Reports, 733, 1, doi: 10.1016/j.physrep.2017.12.002 arXiv:1412.6980. https://arxiv.org/abs/1412.6980
18 Hong et al.

Kourkchi, E., Courtois, H. M., Graziani, R., et al. 2020, Ronneberger, O., Fischer, P., & Brox, T. 2015, arXiv
The Astronomical Journal, 159, 67, e-prints, arXiv:1505.04597.
doi: 10.3847/1538-3881/ab620e https://arxiv.org/abs/1505.04597
Larson, D., Dunkley, J., Hinshaw, G., et al. 2011, The Rubin, V. C., & Ford, W. Kent, J. 1970, The Astrophysical
Astrophysical Journal Supplement Series, 192, 16, Journal, 159, 379, doi: 10.1086/150317
doi: 10.1088/0067-0049/192/2/16 Schaye, J., Crain, R. A., Bower, R. G., et al. 2015, Monthly
Notices of the Royal Astronomical Society, 446, 521,
Lavaux, G., & Hudson, M. J. 2011, Monthly Notices of the
doi: 10.1093/mnras/stu2058
Royal Astronomical Society, 416, 2840,
Shandera, S., Jeong, D., & Grasshorn Gebhardt, H. S.
doi: 10.1111/j.1365-2966.2011.19233.x
2018, Physical Review Letter, 120, 241102,
Lavaux, G., & Jasche, J. 2016, Monthly Notices of the doi: 10.1103/PhysRevLett.120.241102
Royal Astronomical Society, 455, 3169, Shirasaki, M., Yoshida, N., & Ikeda, S. 2019, Physical
doi: 10.1093/mnras/stv2499 Review D, 100, 043527,
Libeskind, N. I., Yepes, G., Knebe, A., et al. 2010, Monthly doi: 10.1103/PhysRevD.100.043527
Notices of the Royal Astronomical Society, 401, 1889, Smith, L. N. 2015, arXiv e-prints, arXiv:1506.01186.
doi: 10.1111/j.1365-2966.2009.15766.x https://arxiv.org/abs/1506.01186
Licquia, T. C., & Newman, J. A. 2015, The Astrophysical Springel, V., Pakmor, R., Pillepich, A., et al. 2018, Monthly
Notices of the Royal Astronomical Society, 475, 676,
Journal, 806, 96, doi: 10.1088/0004-637X/806/1/96
doi: 10.1093/mnras/stx3304
Marinacci, F., Vogelsberger, M., Pakmor, R., et al. 2018,
Springob, C. M., Magoulas, C., Colless, M., et al. 2014,
Monthly Notices of the Royal Astronomical Society, 480,
Monthly Notices of the Royal Astronomical Society, 445,
5113, doi: 10.1093/mnras/sty2206 2677, doi: 10.1093/mnras/stu1743
Milletari, F., Navab, N., & Ahmadi, S.-A. 2016, arXiv Tröster, T., Camera, S., Fornasa, M., et al. 2017, Monthly
e-prints, arXiv:1606.04797. Notices of the Royal Astronomical Society, 467, 2706,
https://arxiv.org/abs/1606.04797 doi: 10.1093/mnras/stx365
Modi, C., Feng, Y., & Seljak, U. 2018, Journal of Tully, R. B., & Courtois, H. M. 2012, The Astrophysical
Cosmology and Astroparticle Physics, 2018, 028, Journal, 749, 78, doi: 10.1088/0004-637X/749/1/78
doi: 10.1088/1475-7516/2018/10/028 Tully, R. B., Courtois, H. M., & Sorce, J. G. 2016, The
Astronomical Journal, 152, 50,
Naiman, J. P., Pillepich, A., Springel, V., et al. 2018,
doi: 10.3847/0004-6256/152/2/50
Monthly Notices of the Royal Astronomical Society, 477,
Tully, R. B., & Fisher, J. R. 1987, Atlas of Nearby Galaxies
1206, doi: 10.1093/mnras/sty618
(Cambridge University Press)
Nelson, D., Pillepich, A., Springel, V., et al. 2018, Monthly Tully, R. B., Shaya, E. J., Karachentsev, I. D., et al. 2008,
Notices of the Royal Astronomical Society, 475, 624, The Astrophysical Journal, 676, 184, doi: 10.1086/527428
doi: 10.1093/mnras/stx3040 Tully, R. B., Courtois, H. M., Dolphin, A. E., et al. 2013,
Nelson, D., Springel, V., Pillepich, A., et al. 2019, The Astronomical Journal, 146, 86,
Computational Astrophysics and Cosmology, 6, 2, doi: 10.1088/0004-6256/146/4/86
doi: 10.1186/s40668-019-0028-x Van Der Walt, S., Colbert, S. C., & Varoquaux, G. 2011,
Paturel, G., Petit, C., Prugniel, P., et al. 2003, Astronomy Computing in science & engineering, 13, 22
Vannerom, D. 2019, Proc. Sci., DIS2019, 111,
& Astrophysics, 412, 45, doi: 10.1051/0004-6361:20031411
doi: 10.22323/1.352.0111
Pawlowski, M. S., Pflamm-Altenburg, J., & Kroupa, P.
Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020,
2012, Monthly Notices of the Royal Astronomical
Nature Methods, 17, 261, doi: 10.1038/s41592-019-0686-2
Society, 423, 1109, doi: 10.1111/j.1365-2966.2012.20937.x
Wes McKinney. 2010, in Proceedings of the 9th Python in
Pillepich, A., Nelson, D., Hernquist, L., et al. 2018, Science Conference, ed. Stéfan van der Walt & Jarrod
Monthly Notices of the Royal Astronomical Society, 475, Millman, 56 – 61, doi: 10.25080/Majora-92bf1922-00a
648, doi: 10.1093/mnras/stx3112 Wilman, D. J., & Erwin, P. 2012, The Astrophysical
Planck Collaboration, Aghanim, N., Akrami, Y., et al. Journal, 746, 160, doi: 10.1088/0004-637X/746/2/160
2018, arXiv e-prints, arXiv:1807.06209. Zonca, A., Singer, L., Lenz, D., et al. 2019, J. Open Sour.
https://arxiv.org/abs/1807.06209 Soft., 4, 1298, doi: 10.21105/joss.01298
Zwicky, F. 1933, Helv. Phys. Acta, 6, 110

You might also like