Prepared for submission to JCAP

Neural Networks for cosmological

model selection and feature
importance using Cosmic
arXiv:2410.05209v1 [astro-ph.CO] 7 Oct 2024

Microwave Background data

I. Ocampo,a,1 G. Cañas-Herrera,b S. Nesserisa

a Instituto
de Física Teórica UAM-CSIC, Universidad Autonóma de Madrid, Cantoblanco,
28049 Madrid, Spain
b ESTEC - European Space Agency, Keplerlaan 1, 2201 AZ Noordwijk, The Netherlands

E-mail: [email protected], [email protected],

[email protected]

Abstract. The measurements of the temperature and polarisation anisotropies of the Cos-
mic Microwave Background (CMB) by the ESA Planck mission have strongly supported the
current concordance model of cosmology. However, the latest cosmological data release from
ESA Planck mission still has a powerful potential to test new data science algorithms and in-
ference techniques. In this paper, we use advanced Machine Learning (ML) algorithms, such
as Neural Networks (NNs), to discern among different underlying cosmological models at the
angular power spectra level, using both temperature and polarisation Planck 18 data. We
test two different models beyond ΛCDM: a modified gravity model: the Hu-Sawicki model,
and an alternative inflationary model: a feature-template in the primordial power spectrum.
Furthermore, we also implemented an interpretability method based on SHAP values to evalu-
ate the learning process and identify the most relevant elements that drive our architecture to
certain outcomes. We find that our NN is able to distinguish between different angular power
spectra successfully for both alternative models and ΛCDM. We conclude by explaining how
archival scientific data has still a strong potential to test novel data science algorithms that
are interesting for the next generation of cosmological experiments.
Contents

1 Introduction 1

2 Methodology 4
2.1 Theoretical predictions in beyond-ΛCDM models 6
2.1.1 Deviations from general relativity 6
2.1.2 Search for primordial features 6
2.2 Machine Learning pipeline using Neural Networks 8
2.2.1 Simulated datasets 8
2.2.2 Neural Network architecture 9

3 Results and discussion 12

3.1 Deviations from general relativity: Hu-Sawicki model 12
3.2 Search for primordial features: linearly-spaced template 13
3.3 ML interpretability and assessment tests 14

4 Conclusions 17

1 Introduction

The Cosmic Microwave Background (CMB) is a remnant radiation from after the Hot Big
Bang, and its detailed observations are crucial to the era of precision cosmology [1]. The
CMB consists of photons that were emitted around 380 000 years after the Big Bang, at a
time when the Universe had cooled enough for neutral atoms to form (at a redshift about
z ∼ 1100). Prior to this period, the Universe was a hot and dense plasma, where photons
were continuously scattering with electrons. For this reason, the CMB represented the photon-
electron last scattering surface of the moment when both types of particles were coupled to
each other. It contains important information about the underlying cosmological model that
best explains the latest observations of our Universe [2, 3].
The Planck satellite, launched by the European Space Agency in 2009 [4–6], is a ref-
erence in modern cosmology thanks to its exquisite measurements of the temperature and
polarisation anisotropies of the CMB [7]. The CMB radiation is highly isotropic with a tem-
perature of approximately 2.7 K, as first studied by the COBE satellite [8], and it is predicted
to present small anisotropies showing under and overdensed regions in the primordial uni-
verse. In fact, the NASA WMAP satellite [9, 10], which followed on COBE, and the ESA
Planck satellite [11], measured the temperature and polarization anisotropies to an excep-
tionally precise level. The discovery of these anisotropies presented in the CMB introduced
a revolutionary approach to understand how our Universe works, and how the initial seeds
that populated the very early Universe evolved to form the current Large-Scale Structure
(LSS) we observed in the Universe. In general, we study the underlying cosmological model
by comparing observable predictions, such as the angular power spectrum decomposition of
the temperature and polarisation anisotropies, against real CMB data. These observations
enabled to impose stringent constraints on the underlying physical model, called the stan-
dard cosmological concordance model, or simply, the ΛCDM (with the best fit values in table
Table 1). This model assumes that the Universe’s accelerated expansion can be explained in

–1–
 Parameter Description Planck alone
Ωb h2 Baryon density parameter 0.022383
Ωc h2 Cold dark matter density parameter 0.12011
100θMC Angular size of the sound horizon 1.040909
τ Optical depth to reionization 0.0543
ln 1010 As Amplitude of the primordial scalar perturbations 3.0448

ns Scalar spectral index 0.96605

H0 km s−1 Mpc−1 Hubble constant 67.32
 

ΩΛ Dark energy density parameter 0.6842

Ωm Matter density parameter 0.3158
σ8 Matter fluctuation amplitude on 8 h−1 Mpc scale 0.8120

Table 1: The ΛCDM best fit 6 cosmological parameters from Planck CMB temperature and
polarization power spectra. The table is divided in the main sampled and derived parameters,
like H0 or σ8 .

terms of a Cosmological constant in the dynamic equations (Λ, popularly associated to Dark
Energy), that the main energy density component is Cold Dark Matter (CDM), and that the
primordial density perturbations in the early Universe were almost Gaussianly-distributed,
compatible with the predictions of the vanilla single-field inflation model. On top, CMB data
has given powerful insights on alternative inflationary models [7], the nature of dark matter
and dark energy [5, 12, 13], and the homogeneity of our Universe [14].
Despite its success in explaining observational phenomena, the ΛCDM model has faced
some challenges and is currently the focus of intense research. There is a lack of understand-
ing of its two main components (Dark Matter and Dark Energy) and also, the precision level
we have reached in parameter extraction, could imply potential tensions in our current un-
derstanding of the Universe [15–17], which include the cosmic dipoles problem [18], the age
of the Universe [19], the lithium problem [20], etc1 . Perhaps the most debated problem is
the so-called Hubble tension, which arises during the first release of cosmological parameters
from the Planck Collaboration [4], from the discrepancy of the Hubble constant found by
CMB early Universe observations: H0 = 67.4 ± 0.5 km s−1 Mpc−1 and the value obtained by
late-time surveys like SH0ES: H0 = 73.04 ± 1.04 km s−1 Mpc−1 [22, 23].
Understanding the late-time Universe’s acceleration and its relation with the Hubble
tension is a current hot discussion topic in cosmology [24, 25]. The discrepancies on its value
may arise from observational systematic effects, statistical fluctuations, or new physics beyond
the ΛCDM model. This is one of the reasons that, despite the ΛCDM model being a very
good statistical fit that explains most of the observational phenomena, there is still interest
in discovering few extension models that can explain the unresolved cosmological late-time
accelerated expansion but yet recover the ΛCDM expansion history and pass the Solar System
tests [26, 27]. Many viable Modified Gravity (MG) theories feature “screening mechanisms´´
that cause deviations from General Relativity (GR) to switch off on small scales, leaving them
with significantly different predictions from GR only over cosmological distances [28].
Within the MG framework, the f (R) family of theories is an interesting alternative
beyond ΛCDM, since it is the most natural extension performed in the Einstein-Hilbert
action, at the level of the Ricci scalar, i.e. R → f (R) ≃ R + mR2 [29]. Some studies on
1
For an extensive review, see [21].

–2–
cosmological constraints on these theories can be found in [30–32]. A popular viable model
within these theories that is in agreement with the ΛCDM expansion history and that passes
the Solar System tests is the Hu-Sawicki (HS) model [33]. In this case, the corresponding
expansion of f (R) takes the form,

H02 R̄02
f (R) = −6ΩDE,0 + |fR0 | + ..., (1.1)
c2 R
where fR0 = df (R)/ dR|z=0 and for values of |fR0 | ≪ 1 the ΛCDM expansion history is well
recovered. In this work, we choose the value of |fR0 | = 10−6 , so that we are still in agreement
with observations while we perform a comparison of models [34].
Furthermore, the ΛCDM model also fails to give an explanation about the nature of the
origin of the Universe. The anisotropies of the CMB indicates that the primordial density
perturbations that populated the very early Universe should have been almost Gaussianly
distributed, explaining the observed homogeneity and isotropy of the cosmos. Inflation is
widely regarded as the mechanism responsible for the origin of these primordial density per-
turbations and, by construction, solves the horizon and flatness problems. This exponentially
rapid expansion phase took place fractions of a second after the Big Bang provides a solid
framework for the generation of primordial density fluctuations that raised the current LSS.
These fluctuations are imprinted on the primordial power spectrum (PPS), which describes
the distribution of these density perturbations across different scales. In the simplest infla-
tionary model, vanilla canonical single-field inflation, this power spectrum is usually defined
as:  ns −1
2π 2 k
PR,0 (k) = 3 PR,0 (k) = As , (1.2)
k k⋆
where ns and As are the spectral index and the amplitude of the comoving curvature pertur-
bations, respectively, and k⋆ = 0.05 Mpc−1 is the pivot scale.
In this sense, Equation 2.6 is compatible with the assumptions of the ΛCDM, setting a
phenomenological parametrisation of an almost-scale invariant power spectrum. However, it
lacks an intrinsic mechanism to explore further inflationary scenarios or further alternatives,
making the study of deviations from this power-law a hot topic in cosmology [35].
Over the past years, an explosion of data has significantly contributed to numerous
studies for testing deviations from ΛCDM. These works have been carried out with data
from late time Universe, like BOSS [36], DES [37], KiDS [38] and early Universe with the
characterization of CMB flcutuations with WMAP [10], Planck [6] and ACT [39]. In the
future, it is expected that the most precise constraints of the cosmological model parameters
will come from stage IV surveys such as the ESA Euclid mission [40–42], LSST [43] and DESI
[44], which are promising to constrain extensions beyond the ΛCDM model.
This explosion of data collection, beyond allowing an increase in quantity and quality,
will pose questions on how to efficiently analyse the data and how to adapt computational
pipelines. This is one of the main reasons why Machine Learning (ML) has become more
popular in cosmology in the past years. The primary benefits of ML techniques lay in its ability
to uncover patterns, insights, and knowledge from large datasets without explicit instructions,
and to adapt and enhance its performance independently. Therefore, ML methods can fully
unlock the potential of multi-probe cosmology by effectively improving systematic effects and
optimally integrating information from various surveys [45]. It facilitates the detection and
classification of cosmological sources as well as the extraction of information from images [46,
47]. ML methods have also been introduced to cosmological simulations for robust predictions

–3–
(see [48]). Additionally, some work has been done in the context of cosmological parameter
inference [49–51] and for testing models beyond ΛCDM in the context of galaxy clustering [34]
and weak lensing [52]. Furthermore, some studies using CMB data and ML were performed,
for example for lensing reconstruction [53] and deep learning of CMB radiation temperature
[54].
While the official operations of the ESA Planck mission came to an end, after showing
the capacity of the CMB temperature and polarisation anisotropies in constraining the un-
derlying cosmological model, it can still be leveraged to provide stringent tests for theories of
fundamental physics beyond the ΛCDM. In the context of the late-time Universe we can test
MG models related to dark energy, like the Hu-Sawicki model, and in the case of the early
Universe we can use the data it to search for deviations in the primordial power spectrum
power-law (the so-called features). Therefore, the aim of this work is to test both cases using
Planck 18 CMB angular power spectrum, to perform ML- based model selection. For this, we
train and test a Neural Network with first, HS simulated data vs ΛCDM simulated data and
second, feature models simulated data vs ΛCDM, both cases taking into account the Planck
18 data uncertainties.
The layout of our paper is the following. In section 2, we introduce the two beyond-
ΛCDM cases we want to test, the Hu-Sawicki model and the linear feature model in the
primordial power spectrum (subsection 2.1), and we explain how we simulated our training
data and created our neural network architecture (subsection 2.2). In section 3, we show
the results of the analysis and comment on the main characteristics found: in subsection 3.1
and subsection 3.2 we discuss the results from the analysis of the Hu-Sawicki model and the
primordial feature template respectively, and, finally, we discuss the interpretability of our
machine learning pipeline in subsection 3.3. Finally, in section 4, we summarise our findings,
explain the main lessons learnt and draw possible future work to further exploit our Machine
Learning pipeline.

2 Methodology

The focus of this study is the development and performance evaluation of a robust data
analysis pipeline for machine learning-based model selection, in order to study two cases,
(1) deviations from general relativity within the framework of the Hu-Sawicki model and
(2) searching for primordial features in the primordial power spectrum. To achieve this, we
implemented an architecture based on Neural Networks to work with Planck 18 data at the
level of the CMB angular power spectra Cℓ . The angular power spectra is given by [55, 56]
Z
2
CℓXY = k 2 dkP (k)∆X Y
ℓ (k)∆ℓ (k), (2.1)
π

where XY stands for the TT, TE and EE polarisation patters, PR (k) is the power-law pri-
mordial spectrum given by Equation 2.6, and ∆X ℓ (k) is the transfer function that is related
to the evolution of primordial density fluctuations from their primordial state to the time of
recombination and beyond, and ℓ is the multipole moment [2, 56]. We can also rewrite the
angular power spectra in the common DXY definition:

ℓ(ℓ + 1)CℓXY
DℓXY = . (2.2)
2π

–4–
 6000
CDM
5000 Planck 18
( + 1)C TT/2 4000
3000
2000
TT =

1000
0
100 500 1000 1500 2000
TT

0
100
500 1000 1500 2000

100 CDM
Planck 18
( + 1)C TE/2

50
0
50
TE =

100
150
10 250 500 750 1000 1250 1500 1750
TE

0
10
250 500 750 1000 1250 1500 1750

40 CDM
Planck 18
30
( + 1)C EE/2

20
10
0
EE =

10
20
25 250 500 750 1000 1250 1500 1750
EE

0
25
250 500 750 1000 1250 1500 1750

Figure 1: Top panels: CMB angular power spectra Planck 18 binned temperature TT,
polarisation EE, and cross-correlations TE anisotropies (orange dots, with error bars), jointly
with the best theory prediction from ΛCDM (solid blue line), calculated with CLASS. Bottom
panels: difference with respect to the best fit for all the different components.

–5–
 The transfer function ∆Xℓ (k) is usually calculated using so-called Boltzmann solvers
CAMB [57] or CLASS [58]. An important aspect of Equation 2.1 is that any late-time Universe
or modified gravity model is reflected in the transfer function, subsequently affecting the
angular power spectra, whereas any deviation originating from the early Universe physics
impacts the power-law primordial power spectrum PR (k), which also induces changes in the
angular angular power spectra. This interplay makes it an ideal framework for testing both
scenarios.
We are interested in studying the statistical properties of the CMB anisotropies, charac-
terized by the temperature (TT mode), polarization (EE mode) and cross (TE) angular power
spectra, since the position and amplitude of the peaks and dips of these spectra are sensitive
to the assumed underlying cosmological model. See, for example, Figure 1, which shows the
Planck 18 angular power spectrum binned data2 compared to the best fit prediction of the
ΛCDM model. For every computation of the ΛCDM baseline model, we chose the fiducial
cosmological parameters according to Planck 18 cosmological results [6], as noted in Table 1,
and the software CLASS.

2.1 Theoretical predictions in beyond-ΛCDM models

2.1.1 Deviations from general relativity
The scenario to test for deviations from GR that we will stick to is the HS f (R) model, having
an effect on the transfer function that at the same time has an impact on the angular power
spectrum Equation 2.1 [59]. To see this, we consider a modification to the Einstein-Hilbert
action of the form [33]:
4 √
Z  
1
S = d x −g f (R) + Lm , (2.3)
2κ
where R is the Ricci scalar, Lm is the matter Lagrangian and κ = 8πGN is a constant (with
GN being the Newton’s constant). By varying the action and taking some considerations,
like the sub-horizon and quasi-static approximations, we get to the following second order
differential equation [60]:
δ̈ + 2H δ̇ − 4πGeff ρδ = 0, (2.4)
where the dot represents the derivative with respect to the cosmic time t and,
2 ′
GN 1 + 4 ka2 FF
Geff = . (2.5)
F 1 + 3 k22 F ′
a F

is the effective gravitational constant that accounts for the effect of modified gravity [60, 61].
Here F (R) ≡ f ′ (R), the prime denotes derivative with respect to R, and a is the scale factor.
Therefore, in this context, the only part affected by f (R) is the transfer function. See [62]
for a reference and [63–65] for current constraints on this model.

2.1.2 Search for primordial features

All cosmological observations so far align with the assumption of adiabatic, Gaussian, and
nearly scale-invariant initial conditions. These results strongly support a specific inflation-
ary scenario known as slow-roll inflation. This pattern is broadly accepted as matching the
predictions of the simplest inflationary single-field canonical model [66, 67]. However, since
2
The data can be found at the Planck Legacy Archive at http://pla.esac.esa.int/pla.

–6–
our observations only capture perturbations, deducing the background that produced them
reveals that various underlying theories can lead to the same set of cosmological observ-
ables. If there are any features in the primordial power spectrum (deviations from the nearly
scale-invariant power spectrum), they would provide a distinctive and revealing insight into
the fundamental theory behind the mechanism that created these initial fluctuations [35].
Primordial features, ∆PR /PR,0 , are parametrised as small deviations from the power-law
primordial power spectrum introduced in Equation 2.6 as:
 
∆PR
PR (k) = PR,0 (k) 1 + (k) , (2.6)
PR,0

where PR,0 is given by Equation 2.6. There are several theories that predict different SHAPes
for features (localised, linear or logarithmic spaced ones). In this paper we study a toy-model
feature template with oscillations linearly spaced in Fourier space with a constant amplitude,
superimposed on the power-law (see Figure 2):
 
∆PR k
= Alin sin ωlin + ϕ , (2.7)
PR,0 k⋆

where Alin is the amplitude of the feature, ϕ is some arbitrary phase and ωlin is the frequency
of oscillations. The fiducial values of the parameters that define the linear oscillations Θlin
are given by [68],
Θlin = {Alin = 0.01, ωlin = 10, ϕlin = 0} . (2.8)

1.15
Linearly-spaced template
1.1

1.05
P (k)

0.95

0.9
3 2 1 0
10 10 10 10
k [Mpc 1]
Figure 2: Linearly-spaced primordial feature ∆PR on the power spectrum of curvature
perturbations, given the fiducial values at Equation 2.8, as a function of scale, in logarithmic
space.

The discovery of any additional inflationary signals hidden in the data could significantly
change our understanding of the early stages of the Universe. While, several feature templates
have been tested against Planck 18 data [7], developing an alternative pipeline to test features

–7–
could offer a way to test extensive photometric and spectroscopic surveys, which offer a
complementary perspective on large-scale structures (LSS). This enhances the sensitivity to
high-frequency signals and complements CMB measurements.

2.2 Machine Learning pipeline using Neural Networks

The Planck mission provided detailed measurements of the anisotropies of the CMB, allowing
to perform cosmological inference on dark energy and modified gravity [5] and inflation [4]
models, as well as putting the tightest constraints on cosmological parameters [6] by sampling
the posterior distribution of the parameters of interest using Markov Chain Monte Carlo
(MCMC) exploration of the likelihood. Instead, this work has the objective to test GR
deviations from ΛCDM and search for primordial features leveraging the power of Neural
Networks (NN). NNs are known to be a powerful tool consisting of many connected units
known as neurons that produce a sequence of real-valued activations, and are widely used
in classification tasks. Within these algorithms, deep neural networks are one of the most
popular and are characterized for having fully connected (dense) layers, i.e. all possible
connections layer-to-layer are present [69]. Deep NN can enhance the accuracy of model
predictions and improve our understanding of the physical processes behind certain patterns.
This is an important task in cosmology, since distinguishing between models could potentially
lead to some insights into the nature of the evolution of the Universe, dark matter and dark
energy, and could set a preliminary starting point for an exhaustive (classical) statistical
analysis on a later time. The analysis pipeline and datasets are publicly available on GitHub3
and Zenodo4 , respectively.

2.2.1 Simulated datasets

To train and test our NN, we simulated two types of datasets:5

• Deviations from general relativity (Hu-Sawicki model): We generated the dataset

using MGCLASS6 [70], a modified version of the Boltzmann solver CLASS, within the frame-
work of the HS model. We simulated 1000 samples (since the scalability of the code does
not compensate the performance of the NN) for ΛCDM baseline model (see Table 1)
varying Ωcdm within the range [0.1, 0.15], and 1000 for the f (R) model with grids of
values |fR0 | ∈ [10−6 , 5 × 10−6 ], for three components CℓT T , CℓT E and CℓEE . For this
case, Planck 18 uncertainties derived from the corresponding covariance matrix, are
subsequently incorporated into the data. Finally to study the point in which the NN
performance breaks down for this section, we also generated data for |fR0 | for different
ranges spanning [10−4 , 5 × 10−4 ], [10−6 , 5 × 10−6 ], [10−7 , 5 × 10−7 ] and [10−8 , 5 × 10−8 ].

• Search for primordial features (linearly-spaced template): We generated a syn-

thetic dataset using a modified version of the Boltzmann solver code CLASS [58] by
introducing the target linearly-spaced feature in the primordial power spectrum (Equa-
tion 2.7). We used the Planck best fit cosmological parameters Table 1 to generate 500
samples (once again, because the scalability of the code does not compensate the perfor-
mance of the NN), of the angular power spectrum for ΛCDM varying Ωcdm ∈ [0.05, 0.15],
3
https://github.com/IndiraOcampo/CMB_ML_based_model_selection.git
4
Primordial feature template (10.5281/zenodo.13829664), and Hu-Sawicki (10.5281/zenodo.13828966)
5
Note that we use the term datasets to refer to the synthetic self-generated data used to train the neural
network (using Planck 18 uncertainties), and it should not be mistaken with the real Planck 18 data.
6
https://gitlab.com/zizgitlab/mgclass--ii/s

–8–
 and 500 for the linearly-spaced primordial feature, varying Ωcdm in the same range and
Alin ∈ [10−2 , 5 × 10−2 ]. We followed this procedure for each of the three components:
CℓT T , CℓT E and CℓEE and added the uncertainties from the covariance matrix set out of
the Planck errors. We also aimed to test the point in which the NN performance breaks
down, so we also generated data for Alin for different ranges between [10−3 , 5 × 10−3 ],
[10−4 , 5 × 10−4 ], [10−5 , 5 × 10−5 ] and [10−6 , 5 × 10−6 ].

2.2.2 Neural Network architecture

Figure 3: NN architecture for the TT case with 83 neurons in the input layer (that accounts
for the 83 data points of each CℓT T realization). We set the batch size to the default 32 value,
followed by the implementation of a fully connected layer of 32 neurons, a dropout layer and
an output layer with one neuron.

Different architectures can influence how effectively a NN can capture underlying pat-
terns and features within the data. In this paper, we chose deep neural networks, because of
their potential in classification tasks. We developed 4 different deep NN: one for each CMB
angular power spectra case (TT, TE, EE), and a final one for the joint case (TT + TE +
EE). A generalized scheme of the implemented NN’s architecture is shown in Figure 3, where
the units (neurons) in the input layer depended on the number of data points of each case:
for CℓT T − 83 units, for CℓT E − 66 units, for CℓEE − 66 units and the last one for the combined
analysis CℓT T + CℓT E + CℓEE − 215 units. The units correspond to the number of data points
present in Planck 18 binned data.

The efficiency of training a deep NN heavily relies on the quality of the input data.
Therefore, it is essential to assess, clean, and process the data to maximize the NN’s ability

–9–
 4 CDM
Feature template
3
TT
2

1
Normalized

0 500 1000 1500 2000 2500

Figure 4: Illustration of how the Z-score normalization for the CℓT T CMB power spectrum
maximize the differences between ΛCDM and extension models (in this case, the linearly-
spaced feature template as in Equation 2.7). The Z-score normalization not only re-scales the
mean of the CℓT T but also highlights the difference in phase for the CℓT T values as a function
of ℓ.

to learn key features. Common approaches for achieving this are statistical normalization
techniques. In this case, we applied the so-called Z-score normalization for the data pre-
processing, which consists on measuring how many standard deviations σ a data point is
from the mean of the dataset µ, and use this reference setup to effectively re-scale the whole
dataset so that µ = 0 and σ = 1 [71]. To exemplify this in our analysis, we show the impact of
using Z-score normalization in one of our datasets; in particular, on the CℓT T component for
the feature model case (see Figure 4). The differences on CℓT T after normalization between
ΛCDM and the linearly-spaced feature template are more evident, allowing the deep NN to
easily identify possible features for further classification and identification. Similar results
are found for the HS model. We performed the same Z-score normalization for the pre-
processing of all 8 architectures (4 for the HS model and 4 for the primordial feature model).
Furthermore, we have assessed the impact of using the Z-score normalization on the Dℓ (see
Equation 2.2) instead of Cℓ to train the NN, not finding significant differences in the output
and therefore, re-assuring the stability of the NN architecture.
After completing the data pre-processing steps, we divided the 1500 data realizations
(with 50% corresponding to ΛCDM and 50% to either HS or primordial features, depending
on the scenario) into two sets: 70% for training and 30% for testing. We set the batch
size to 32, and incorporated a hidden dense layer with 32 units with a ReLU activation
function [72]. We included a dropout layer (with a dropout rate of 0.2) as regularisation
technique to prevent over-fitting [73]. Finally, we added the last dense layer with one unit
and a sigmoid activation function that enhances the classification score for the model: 0−
ΛCDM and 1− HS or primordial feature. The models were compiled with an Adam optimizer
(learning_rate = 0.0002), a binary cross entropy loss function, and trained over a 1000 epochs.

– 10 –
See Figure 3 for a schematic view of one of the NN’s architecture.

Hu-Sawicki model Linearly-spaced template

6000
CDM CDM
5000 fR0=1.e-7 5000 Alin=1.e-2
4000
( + 1)C TT/2

4000
3000 3000

TT
2000
TT =

2000
1000 1000
0 0
25
250 500 750 1000 1250 1500 1750 2000
500 1000 1500 2000 2500

0.02
TT

TT
0.00 25
0.02 50
250 500 750 1000 1250 1500 1750 2000 500 1000 1500 2000 2500

CDM
100 CDM
fR0=1.e-7 100 Alin=1.e-2
50
( + 1)C TE/2

50
0 0
TE

50
TE =

50
100 100
150 1e 5
2 250 500

Multipole moment
750 1000 1250 1500 1750 2000

2 250 500 750 1000 1250 1500 1750 2000

1
TE

TE

0
0
1 2
250 500 750 1000 1250 1500 1750 2000 250 500 750 1000 1250 1500 1750 2000

CDM
40 fR0=1.e-7 40 CDM
Alin=1.e-2
( + 1)C EE/2

30 30
EE

20 20
EE =

10 10

0 1e 10 0
1.0
Multipole moment
250 500 750 1000 1250 1500 1750 2000

0.5
250 500 750 1000 1250 1500 1750 2000

0.5
EE

EE

0.0 0.0
0.5 0.5
250 500 750 1000 1250 1500 1750 2000 250 500 750 1000 1250 1500 1750 2000

Figure 5: Left: fR0 effect on the TT, TE and EE components. Right: Feature template on
the primordial power spectrum, this is reflected on the effect of the Alin parameter size on
the TT, TE and EE components.

– 11 –
3 Results and discussion

In this section, we present the main results obtained with our pipeline and the outcome
of the learning assessment with ML interpretability. The classification performance of our
NN’s architecture is available in Table 2 for the three components CℓT T , CℓT E , CℓEE and its
combination CℓT T + CℓT E + CℓEE , for both the two cases under study (deviations from general
relativity and features in the primordial power spectrum). We used ranges of parameters
that encompassed the fiducial values of both models as Alin = 10−2 and |fR0 | = 10−6 (see
Figure 5). From these results, we can see that the NNs perform better in detecting the
alternative initial conditions in the context of the early Universe physics than the deviations
from GR in the late-time Universe.

Hu-Sawicki model linearly-spaced feature

CMB components
Correct Wrong Correct Wrong
CℓT T 0.93 0.07 1 0
CℓT E 0.47 0.53 1 0
CℓEE 0.48 0.52 1 0
CℓT T + CℓT E + CℓEE 0.88 0.12 1 0

Table 2: Performance of Neural Networks classification when trained and tested on each
component of the temperature and polarization angular Power Spectrum, and its combined
analysis. Two models were studied: for the early Universe case, a linearly-spaced template
to search for primordial features in the primordial power spectrum (with a range of values
of Alin ∈ [10−2 , 5 × 10−2 ]), and for the late-time Universe, the Hu-Sawicki model to test for
deviations from ΛCDM (with a range of values of |fR0 | ∈ [10−6 , 5 × 10−6 ]).

3.1 Deviations from general relativity: Hu-Sawicki model

In the context of MG, the NN performance in detecting CℓT T coming from ΛCDM or HS is
0.93 (see Table 2), compared to the cases of the other components CℓT E and CℓEE , where the
performance drops to 0.47 and 0.48. When studying the case in which we combined all the
data, we concluded that the NN is mostly learning to differentiate between models mainly
from the temperature angular power spectrum component.
To understand this, we illustrate in Figure 5 the effect of fR0 on the Cℓ components
of the CMB, where only the CℓT T one is affected by the changes in this parameter. This is
why the performance of the NN to classify between ΛCDM and HS is clearly better in the
temperature angular power spectrum than the other two cases: CℓT E and CℓEE do not show
any deviations in HS with respect to ΛCDM. We also aimed to study the point (if any) in
which the NNs performance breaks down or changes, so we generated datasets for different
ranges of fR0 , and then compared the performance on each of them of them, and also the
combined case CℓT T + CℓT E + CℓEE . This information is available in Figure 6, where we can
notice that for the CℓT T case, the performance breaks down for values of |fR0 | = 10−7 , and
for the other two modes, the NN cannot differentiate between models regardless of the value
of this parameter. This value has proven to be still in agreement with current surveys (see,
for example [74]), while the ΛCDM model is recovered with |fR0 | = 0. In fact, the NN was
able to learn a feature of the HS model: the HS model only shows deviations on the CℓT T
component (see [75] where they verify that this is indeed the case by comparing Boltzmann

– 12 –
 solvers). This blind test strengthens the robustness of our pipeline, as the NN could not
classify the data realization as neither ΛCDM nor HS using CℓT E and/or CℓEE .

1.0 0.53
Correct Prediction
0.52 Wrong Prediction

Performance (Trained on C TE)

Performance (Trained on C TT)

0.8
0.51
0.6
Correct Prediction 0.50
Wrong Prediction
0.4
0.49
0.2 0.48
0.0 0.47
8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0 8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0
Log(fR0) Log(fR0)

0.53
Performance (Trained on C TT + C TE + C EE)
Performance (Trained on C EE)

0.52 0.8
0.51
0.6
Correct Prediction Correct Prediction
0.50 Wrong Prediction Wrong Prediction
0.4
0.49
0.48 0.2

0.47 0.0
8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0 8.0 7.5 7.0 6.5 6.0 5.5 5.0 4.5 4.0
Log(fR0) Log(fR0)

Figure 6: Semi-log plot of fR0 values vs architecture performance for the components (up)
CℓT T , CℓT E (bottom) CℓEE , joint analysis: CℓT T + CℓT E + CℓEE . The yellow region corresponds
to the fiducial value of |fR0 | = 10−6 .

3.2 Search for primordial features: linearly-spaced template

With respect to the case for features in the primordial power spectrum, defined as linearly-
spaced, the NN is able to differentiate between models in all three angular power spectra, CℓT T ,
CℓT E and, CℓEE , and also for the combined case. In Table 2, we notice that the performance
is 1 for all the cases, which is sensible if we see the effect of Alin on the modes in ?? for the
fiducial value of Alin = 10−2 , where clear deviations are shown with respect to the ΛCDM
model. This is expected since primordial features in the power spectrum are expected to
introduce changes on all CMB Cℓ components, as given by Equation 2.1.
For completion, we also analyzed the point in which the NN’s performance breaks down,
and for all the cases there is similar value of Alin = 10−4 , where the correct prediction decreases
to 0.4 for CℓT T and CℓT E and to ≈ 0.6 for CℓEE and the combined case (see Figure 7). In
general, the NN is able to discern between ΛCDM and the linearly-spaced feature model for

– 13 –
fiducial values of Alin . This is a highlight of our pipeline, since the NN is able to classify well
even when the differences between the models are very small.

1.0 Correct Prediction 1.0 Correct Prediction

Wrong Prediction Wrong Prediction

Performance (Trained on C TE)

Performance (Trained on C TT)

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0
Log(Alin) Log(Alin)

1.0 1.0
Performance (Trained on C TT + C TE + C EE)

Correct Prediction Correct Prediction

Wrong Prediction Wrong Prediction
Performance (Trained on C EE)

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0
Log(Alin) Log(Alin)

Figure 7: Semi-log plot of Alin values vs architecture performance for the components (up)
CℓT T , CℓT E (bottom) CℓEE , joint analysis: CℓT T + CℓT E + CℓEE . The cyan region corresponds
to the fiducial value of Alin = 10−2 .

3.3 ML interpretability and assessment tests

Machine Learning (ML) interpretability has become essential in deep learning, as understand-
ing how these ML techniques reach their conclusions is necessary for building trust in their
application across various fields, particularly in analyzing complex datasets [76]. To assess
how the NN learn, we focused our work on the interpretability tool known as SHAP (SHAp-
ley Additive exPlanations)7 . In general, ML interpretability follows two approaches: local
and global explanations. Global explanations provide an overview of the learning’s process
throughout the entire dataset, whereas the local method focuses on studying a region in the
hyperparameter space. SHAP, which can be employed for both local and global explanations,
7
https://github.com/SHAP/SHAP.git

– 14 –
 HS
0.75 3

0.50
(Local explanations summary)

Feature Value (Low to High)

2

0.25
1
SHAP Value

0.00
0
0.25
1
0.50
2
0.75
3 CDM
48 500 1000 1500 2000 2500

1e 10
TT (HS or CDM)
8
Mean Absolute SHAP Value

6
(global explanations)

0
500 1000 1500 2000 2500

Figure 8: Machine Learning explanations for the Hu-Sawicki model. Top: local interpretabil-
ity given by the SHAP values (high value, Hu-Sawicki and low value, ΛCDM). Bottom: global
interpretability by the mean absolute SHAP values that corresponds to each CℓT T .

is a method for explaining the output of a NN architecture by attributing the contribution of

each element to the final prediction, which is known as feature importance 8 .
SHAP is based on cooperative game theory, and helps to understand the most important
input elements (or "players") by computing an adapted version of the so-called Shapley values
[77]. It has a bottom-up approach: for the local explanations, it assesses the contribution
of each input datapoint (in our case, the Cℓ ) to the prediction, by considering all possible
combinations. In this context, Shapley values are calculated by introducing each element x
8
Not to be confused with the feature template of the primordial power spectrum in extensions beyond
single-field inflation; therefore in the body of the text of this work, we will refer to it as "element importance".

– 15 –
(one by one) into a conditional function of the model’s output

fx (S) ≈ E [f (x) | xS ] , (3.1)

which can be understood as the expected prediction given that only the elements in a subset
S are known. The change in the model’s decision at each step is attributed to the newly
introduced element. This process is then averaged across all possible element orderings to
ensure a fair distribution of contributions. Hence, the SHAP values produces attributions
ϕi (f, x) that matches the original model output f (x) [77], given by,
X 1
ϕi (f, x) = [fx (Pim ∪ i) − fx (Pim )] , (3.2)
N!
m∈M

where M is the set of all feature orderings, Pim is the set of all elements that come before
the ith element in ordering m, and N is the number of input elements for the model. For a
revision on this topic refer to [77, 78] and [79]. On the contrary, for the global interpretability
approach, this method calculates the mean absolute SHAP values, giving an overall importance
to each element for the model’s outcome.
In this work, SHAP is used for one of the first times as interpretability tool in cosmological
model selection and interpretability [80, 81]. We employed SHAP for both, local and global
interpretability assessments for the temperature angular power spectrum CℓT T for the two
extension models. The analysis takes approximately 3 hours for each model. For the case
of the modified gravity model, we can see the Hu-Sawicki NN explanations in Figure 8; the
upper panel displays the local interpretability assessment with SHAP values that correspond to
each DℓT T . Higher values mean that their corresponding elements are more likely to contribute
to a Hu-Sawicki outcome, and lower values to a ΛCDM one. In the lower panel, we see the
global interpretability results with the mean absolute SHAP values, that correspond to each
DℓT T . The dots represent one randomly selected realization of the data (either belonging to
the HS model or to ΛCDM). We observe how the first DℓT T values are the critical one for the
NN to learn the main elements to classify either ΛCDM or HS. This is expected as observed
in Figure 5, where the main changes with respect to ΛCDM for the HS model take place at
low multipole values.
Subsequently, in figure Figure 9, we show the explanations for the linearly-spaced pri-
mordial feature template model; as in the previous case, the upper panel displays the local
interpretability part with SHAP values that correspond to each DℓT T . In this case, higher values
show that their corresponding elements are more likely to contribute to a feature-template
model outcome, and lower values to a ΛCDM one. In the same way, the lower panel displays
the global interpretability conclusions, with the mean absolute SHAP values, that correspond
to each DℓT T . Again, the dots represent one randomly selected realization of the data (ei-
ther belonging to the feature-template model or to ΛCDM). In this case, both the local and
global explanations trace properly the changes introduced in the DℓT T due to the primordial
feature template in Figure 2, identifying as critical those ℓ where the corresponding Planck
18 uncertainty is larger, and therefore, more likely to allow for deviations from the power-law
primordial power spectrum. In summary, we notice that the NN are properly learning how
to differentiate models based on the deviations displayed in the lower panels of Figure 5, be-
cause it seems that the peaks of the most important elements in both Figure 8 and Figure 9
coincide with these deviations. We also raise an interesting point by revisiting Table 2, where
we conclude that NN perform better in differentiating the primordial feature template from

– 16 –
 Primordial
4 feature
0.6
3
0.4
(Local explanations summary)

Feature Value (Low to High)

2
0.2
SHAP Value

1
0.0
0
0.2
1
0.4
2
0.6
CDM
48 500 1000 1500 2000 2500

1e 9
TT (Primordial feature or CDM)
1.0
Mean Absolute SHAP Value

0.8
(global explanations)

0.6

0.4

0.2

0.0
500 1000 1500 2000 2500

Figure 9: Machine Learning explanations for the linearly-spaced primordial feature model.
Top: local interpretability given by the SHAP values (high values, primordial feature model
and low ones, ΛCDM). Bottom: global interpretability by the mean absolute SHAP values that
corresponds to each CℓT T .

ΛCDM than the case of Hu-Sawicki. In light of the SHAP values, this is expected because the
primordial feature model displays more numerous important elements than those of the HS
analysis, helping the NN identify the key-elements for classification.

4 Conclusions

In this work, we presented our Machine Learning pipeline designed for model selection using
deep Neural Networks to classify two models beyond the Standard Cosmological Model. Our
first focus was on late-time physics, specifically deviations from General Relativity within the

– 17 –
framework of the Hu-Sawicki f (R) model. The second focus was on early universe hypotheses,
particularly the search for linearly-spaced features in the primordial power spectrum. The
architecture of our deep Neural Network pipeline has been optimized to maximize the learning
of key features while minimizing over-fitting.

To train our pipeline, we generated Planck 2018-like data, including its specifications and
uncertainties for the mentioned cases. We tested its ability to distinguish between data derived
from the Hu-Sawicki model vs ΛCDM and the primordial feature model vs ΛCDM. Our
pipeline comprises a total of eight Neural Networks: four dedicated to the three polarization
modes and one that combines all of them for testing the linearly spaced primordial feature
model, as well as four for evaluating the Hu-Sawicki model. Generally, the architectures
remained consistent, with the primary variation being the number of neurons in the input
layer, which corresponded to the number of binned Planck 2018 angular power spectrum data
points.

The results demonstrate that our pipeline effectively learned the crucial elements neces-
sary to classify the test dataset. Specifically, for the Hu-Sawicki datasets, the Neural Network
achieved a correct prediction rate of 93% for the temperature power spectrum. However, for
polarization and cross-data, the performance dropped to 0.47 and 0.48, respectively, aligning
with theoretical expectations, as the Hu-Sawicki model does not exhibit differences compared
to ΛCDM in the polarization CMB angular power spectra. This outcome indicates that our
Neural Network accurately identified this insensitivity and successfully passed this sanity
check. In the second scenario, we aimed to differentiate between the primordial linearly-
spaced feature model and ΛCDM across temperature, polarization, and cross-data (as well as
a combination of all three). Remarkably, the pipeline achieved a 100% correct prediction rate
across all components, making this kind of ML pipelines highly useful in the community to
assess the potential of a dataset to constrain features in the primordial power spectrum. In
fact, this computationally cheaper approach can be applied before addressing a full Bayesian
inference exercise to sample the probability posterior distributions of the features’ parameters
to actually discern the capacity to detect a possible primordial feature template.

To evaluate and understand how our pipeline is learning to classify amongst models, we
implemented SHAP as an interpretability tool to extract meaningful insights about the main
input elements that drive these Neural Networks to certain outcomes. The application of
this particular interpretability tool is one of the first applications of SHAP to cosmological
model selection. We confirmed that our architectures can differentiate between models by
focusing on the elements of the angular power spectra that reflect the deviations between the
proposed alternatives and ΛCDM, these key elements were identified prior to training. We
conclude with highlighting our ML interpretability results, since feature importance is crucial
for evaluating the learning process and deep learning algorithms do not always make decisions
based on the features or elements we intend for them to learn.

Looking ahead to future work, we see potential for extending our work by applying
the pipeline to classify models using CMB temperature and polarization maps, leveraging
techniques like Convolutional Neural Networks or Graph Neural Networks. Additionally, we
aim to apply the pipeline to other summary statistics from various cosmological observables,
such as cosmic shear in Large-Scale Structure and photometric angular clustering power
spectra.

– 18 –
Acknowledgments
We thank Dr. Stéphane Ilic for interesting discussions about the Planck 18 covariance ma-
trices. IO thanks ESTEC/ESA for the warm hospitality during the execution of this project,
and for support from the ESA Archival Research Visitor Programme. IO thanks to Elena
Donini for useful discussions during the ESA internship. IO and SN acknowledge support
from the research project PID2021-123012NB-C43 and the Spanish Research Agency (Agen-
cia Estatal de Investigación) through the Grant IFT Centro de Excelencia Severo Ochoa No
CEX2020-001007-S, funded by MCIN/AEI/10.13039/501100011033. IO is also supported by
the fellowship LCF/BQ/DI22/11940033 from “la Caixa” Foundation (ID 100010434). GCH
acknowledges support through the ESA research fellowship programme.

References
[1] RB Barreiro. An overview of the current status of cmb observations. Highlights of Spanish
Astrophysics V, pages 93–102, 2010.
[2] Daniel Baumann. Tasi lectures on inflation. arXiv preprint arXiv:0907.5424, 2009.
[3] Wayne Hu. Lecture notes on cmb theory: From nucleosynthesis to recombination. arXiv
preprint arXiv:0802.3688, 2008.
[4] Planck Collaboration, PAR Ade, N Aghanim, C Armitage-Caplan, M Arnaud, M Ashdown,
F Atrio-Barandela, J Aumont, C Baccigalupi, AJ Banday, et al. Planck 2013 results. xxii.
constraints on inflation. 0004-6361, 571:22, 2014.
[5] Peter AR Ade, N Aghanim, M Arnaud, M Ashdown, J Aumont, C Baccigalupi, AJ Banday,
RB Barreiro, Nicola Bartolo, E Battaner, et al. Planck 2015 results-xiv. dark energy and
modified gravity. Astronomy & Astrophysics, 594:A14, 2016.
[6] Nabila Aghanim, Yashar Akrami, Mark Ashdown, Jonathan Aumont, Carlo Baccigalupi, Mario
Ballardini, Anthony J Banday, RB Barreiro, Nicola Bartolo, S Basak, et al. Planck 2018
results-vi. cosmological parameters. Astronomy & Astrophysics, 641:A6, 2020.
[7] Yashar Akrami, Frederico Arroja, M Ashdown, J Aumont, Carlo Baccigalupi, M Ballardini,
Anthony J Banday, RB Barreiro, Nicola Bartolo, S Basak, et al. Planck 2018 results-x.
constraints on inflation. Astronomy & Astrophysics, 641:A10, 2020.
[8] George F. Smoot, Charles L. Bennett, Alan Kogut, Edward L. Wright, Radek Stompor, et al.
Structure in the cobe differential microwave radiometer first-year maps. The Astrophysical
Journal Letters, 396:L1–L5, 1992.
[9] David N Spergel, Licia Verde, Hiranya V Peiris, Eiichiro Komatsu, MR Nolta, Charles L
Bennett, Mark Halpern, Gary Hinshaw, Norman Jarosik, Alan Kogut, et al. First-year
wilkinson microwave anisotropy probe (wmap)* observations: determination of cosmological
parameters. The Astrophysical Journal Supplement Series, 148(1):175, 2003.
[10] Gary Hinshaw, D Larson, Eiichiro Komatsu, David N Spergel, CLaa Bennett, Joanna Dunkley,
MR Nolta, M Halpern, RS Hill, N Odegard, et al. Nine-year wilkinson microwave anisotropy
probe (wmap) observations: cosmological parameter results. The Astrophysical Journal
Supplement Series, 208(2):19, 2013.
[11] Peter AR Ade, Nabila Aghanim, Monique Arnaud, Mark Ashdown, J Aumont, C Baccigalupi,
A Balbi, AJ Banday, RB Barreiro, JG Bartlett, et al. Planck early results. xviii. the power
spectrum of cosmic infrared background anisotropies. Astronomy & Astrophysics, 536:A18,
2011.
[12] Giacomo Galloni, Mario Ballardini, Nicola Bartolo, Alessandro Gruppuso, Luca Pagano, and
Angelo Ricciardone. Unraveling the cmb lack-of-correlation anomaly with the cosmological

– 19 –
 gravitational wave background. Journal of Cosmology and Astroparticle Physics, 2023(10):013,
2023.
[13] Douglas Scott, Dagoberto Contreras, Ali Narimani, and Yin-Zhe Ma. The information content
of cosmic microwave background anisotropies. Journal of Cosmology and Astroparticle Physics,
2016(06):046, 2016.
[14] Y. Akrami et al. Planck 2018 results. VII. Isotropy and Statistics of the CMB. Astron.
Astrophys., 641:A7, 2020.
[15] William Giarè. Cmb anomalies and the hubble tension. arXiv preprint arXiv:2305.16919, 2023.
[16] Jian-Ping Hu and Fa-Yin Wang. Hubble tension: the evidence of new physics. Universe,
9(2):94, 2023.
[17] Elcio Abdalla, Guillermo Franco Abellán, Amin Aboubrahim, Adriano Agnello, Özgür Akarsu,
Yashar Akrami, George Alestas, Daniel Aloni, Luca Amendola, Luis A Anchordoqui, et al.
Cosmology intertwined: A review of the particle physics, astrophysics, and cosmology
associated with the cosmological tensions and anomalies. Journal of High Energy Astrophysics,
34:49–211, 2022.
[18] Chethan Krishnan, Ranjini Mondol, and MM Sheikh-Jabbari. Dipole cosmology: the
copernican paradigm beyond flrw. Journal of Cosmology and Astroparticle Physics,
2023(07):020, 2023.
[19] Licia Verde, Pavlos Protopapas, and Raul Jimenez. Planck and the local universe: Quantifying
the tension. Physics of the Dark Universe, 2(3):166–175, 2013.
[20] Brian D Fields. The primordial lithium problem. Annual Review of Nuclear and Particle
Science, 61(1):47–68, 2011.
[21] Leandros Perivolaropoulos and Foteini Skara. Hubble tension or a transition of the cepheid
snia calibrator parameters? Physical Review D, 104(12):123511, 2021.
[22] Adam G Riess, Wenlong Yuan, Lucas M Macri, Dan Scolnic, Dillon Brout, Stefano Casertano,
David O Jones, Yukei Murakami, Gagandeep S Anand, Louise Breuval, et al. A comprehensive
measurement of the local value of the hubble constant with 1 km s- 1 mpc- 1 uncertainty from
the hubble space telescope and the sh0es team. The Astrophysical journal letters, 934(1):L7,
2022.
[23] Pablo Lemos and Paul Shah. The cosmic microwave background and h_0. arXiv preprint
arXiv:2307.13083, 2023.
[24] Dark Energy Survey Collaboration:, T Abbott, FB Abdalla, J Aleksić, S Allam, A Amara,
D Bacon, E Balbinot, M Banerji, K Bechtol, et al. The dark energy survey: more than dark
energy–an overview. Monthly Notices of the Royal Astronomical Society, 460(2):1270–1299,
2016.
[25] Dark Energy Survey Collaboration: T Abbott et al. The dark energy survey. International
Journal of Modern Physics A, 20(14):3121–3123, 2005.
[26] Jai-chan Hwang and Hyerim Noh. f (r) gravity theory and cmbr constraints. Physics Letters B,
506(1-2):13–19, 2001.
[27] Pedro Bessa, Marcela Campista, and Armando Bernui. Observational constraints on
starobinsky f (r) cosmology from cosmic expansion and structure growth data. The European
Physical Journal C, 82(6):506, 2022.
[28] Alejandro Aviles, Jorge L Cervantes-Cota, and David F Mota. Screenings in modified gravity:
a perturbative approach. Astronomy & Astrophysics, 622:A62, 2019.
[29] Alexei A. Starobinsky. A new type of isotropic cosmological models without singularity.
Physics Letters B, 91(1):99–102, 1980.

– 20 –
[30] Spyros Basilakos, Savvas Nesseris, and Leandros Perivolaropoulos. Observational constraints
on viable f (r) parametrizations with geometrical and dynamical probes. Physical Review
D—Particles, Fields, Gravitation, and Cosmology, 87(12):123529, 2013.
[31] Soumya Jana and Subhendra Mohanty. Constraints on f (r) theories of gravity from gw170817.
Physical Review D, 99(4):044056, 2019.
[32] Rafael C Nunes, Supriya Pan, Emmanuel N Saridakis, and Everton MC Abreu. New
observational constraints on f (r) gravity from cosmic chronometers. Journal of Cosmology and
Astroparticle Physics, 2017(01):005, 2017.
[33] Wayne Hu and Ignacy Sawicki. Models of f (r) cosmic acceleration that evade solar system
tests. Physical Review D—Particles, Fields, Gravitation, and Cosmology, 76(6):064004, 2007.
[34] Indira Ocampo, George Alestas, Savvas Nesseris, and Domenico Sapone. Enhancing
cosmological model selection with interpretable machine learning. arXiv preprint
arXiv:2406.08351, 2024.
[35] Jens Chluba, Jan Hamann, and Subodh P. Patil. Features and new physical scales in
primordial observables: Theory and observation. International Journal of Modern Physics D,
24(10):1530023, June 2015.
[36] Shadab Alam, Metin Ata, Stephen Bailey, Florian Beutler, Dmitry Bizyaev, Jonathan A
Blazek, Adam S Bolton, Joel R Brownstein, Angela Burden, Chia-Hsun Chuang, et al. The
clustering of galaxies in the completed sdss-iii baryon oscillation spectroscopic survey:
cosmological analysis of the dr12 galaxy sample. Monthly Notices of the Royal Astronomical
Society, 470(3):2617–2652, 2017.
[37] Timothy MC Abbott, Michel Aguena, Alex Alarcon, S Allam, O Alves, A Amon,
F Andrade-Oliveira, James Annis, S Avila, D Bacon, et al. Dark energy survey year 3 results:
Cosmological constraints from galaxy clustering and weak lensing. Physical Review D,
105(2):023520, 2022.
[38] Jelte TA de Jong, Gijs A Verdoes Kleijn, Konrad H Kuijken, Edwin A Valentijn, KiDS, and
Astro-WISE Consortiums. The kilo-degree survey. Experimental Astronomy, 35:25–44, 2013.
[39] Simone Aiola, Erminia Calabrese, Loïc Maurin, Sigurd Naess, Benjamin L Schmitt,
Maximilian H Abitbol, Graeme E Addison, Peter AR Ade, David Alonso, Mandana Amiri,
et al. The atacama cosmology telescope: Dr4 maps and cosmological parameters. Journal of
Cosmology and Astroparticle Physics, 2020(12):047, 2020.
[40] Alain Blanchard, S Camera, Carmelita Carbone, VF Cardone, S Casas, Sébastien Clesse, S Ilić,
M Kilbinger, T Kitching, Martin Kunz, et al. Euclid preparation-vii. forecast validation for
euclid cosmological probes. Astronomy & Astrophysics, 642:A191, 2020.
[41] Rene Laureijs, Jérôme Amiaux, S Arduini, J-L Augueres, J Brinchmann, R Cole, M Cropper,
C Dabin, L Duvet, A Ealet, et al. Euclid definition study report. arXiv preprint
arXiv:1110.3193, 2011.
[42] Y Mellier, JA Barroso, A Achúcarro, J Adamek, R Adam, GE Addison, N Aghanim,
M Aguena, V Ajani, Y Akrami, et al. Euclid. i. overview of the euclid mission. arXiv preprint
arXiv:2405.13491, 2024.
[43] Željko Ivezić, Steven M Kahn, J Anthony Tyson, Bob Abel, Emily Acosta, Robyn Allsman,
David Alonso, Yusra AlSayyad, Scott F Anderson, John Andrew, et al. Lsst: from science
drivers to reference design and anticipated data products. The Astrophysical Journal,
873(2):111, 2019.
[44] Behzad Abareshi, J Aguilar, S Ahlen, Shadab Alam, David M Alexander, R Alfarsy, L Allen,
C Allende Prieto, O Alves, J Ameel, et al. Overview of the instrumentation for the dark energy
spectroscopic instrument. The Astronomical Journal, 164(5):207, 2022.

– 21 –
[45] Rémy Joseph, Peter Melchior, and Fred Moolekamp. Joint survey processing: combined
resampling and convolution for galaxy modelling and deblending. arXiv preprint
arXiv:2107.06984, 2021.
[46] Cora Dvorkin, Siddharth Mishra-Sharma, Brian Nord, V Ashley Villar, Camille Avestruz,
Keith Bechtol, Aleksandra Ćiprijanović, Andrew J Connolly, Lehman H Garrison, Gautham
Narayan, et al. Machine learning and cosmology. arXiv preprint arXiv:2203.08056, 2022.
[47] Kana Moriwaki, Takahiro Nishimichi, and Naoki Yoshida. Machine learning for observational
cosmology. Reports on Progress in Physics, 86(7):076901, 2023.
[48] Jonah C Rose, Paul Torrey, Francisco Villaescusa-Navarro, Mariangela Lisanti, Tri Nguyen,
Sandip Roy, Kassidy E Kollmann, Mark Vogelsberger, Francis-Yan Cyr-Racine, Mikhail V
Medvedev, et al. Introducing the dreams project: Dark matter and astrophysics with machine
learning and simulations. arXiv preprint arXiv:2405.00766, 2024.
[49] Koya Murakami, Indira Ocampo, Savvas Nesseris, Atsushi J Nishizawa, and Sachiko
Kuroyanagi. Non-linearity-free prediction of the growth-rate f σ8 using convolutional neural
networks. arXiv preprint arXiv:2305.12812, 2023.
[50] Zhiwei Min, Xu Xiao, Jiacheng Ding, Liang Xiao, Jie Jiang, Donglin Wu, Qiufan Lin, Yin Li,
Yang Wang, Shuai Liu, et al. Deep learning for cosmological parameter inference from dark
matter halo density field. arXiv preprint arXiv:2404.09483, 2024.
[51] Jorge Enrique García-Farieta, Héctor J Hortúa, and Francisco-Shu Kitaura. Bayesian deep
learning for cosmic volumes with modified gravity. Astronomy & Astrophysics, 684:A100, 2024.
[52] Austin Peel, Florian Lalande, Jean-Luc Starck, Valeria Pettorino, Julian Merten, Carlo Giocoli,
Massimo Meneghetti, and Marco Baldi. Distinguishing standard and modified gravity
cosmologies with machine learning. Physical Review D, 100(2):023508, 2019.
[53] Ye-Peng Yan, Guo-Jian Wang, Si-Yu Li, Yang-Jie Yan, and Jun-Qing Xia. Lensing
reconstruction from the cosmic microwave background polarization with machine learning. The
Astrophysical Journal, 952(1):15, 2023.
[54] Mehmet Salti and Evrim Ersin Kangal. Deep learning of cmb radiation temperature. Annals of
Physics, 439:168799, 2022.
[55] Daniel Baumann. Tasi lectures on primordial cosmology. arXiv preprint arXiv:1807.03098,
2018.
[56] Carolyn Judith MacTavish. CMB angular power spectra and cosmological implications from the
2003 LDB flight of the BOOMERANG telescope. University of Toronto, 2006.
[57] Antony Lewis, Anthony Challinor, and Anthony Lasenby. Efficient computation of cosmic
microwave background anisotropies in closed friedmann-robertson-walker models. The
Astrophysical Journal, 538(2):473–476, 2000.
[58] Julien Lesgourgues. The Cosmic Linear Anisotropy Solving System (CLASS) I: Overview.
arXiv e-prints, page arXiv:1104.2932, April 2011.
[59] Raphaël Kou, Calum Murray, and James G Bartlett. Constraining f (r) gravity with
cross-correlation of galaxies and cosmic microwave background lensing. Astronomy &
Astrophysics, 686:A193, 2024.
[60] Shinji Tsujikawa. Matter density perturbations and effective gravitational constant in modified
gravity models of dark energy. Physical Review D—Particles, Fields, Gravitation, and
Cosmology, 76(2):023514, 2007.
[61] Hayato Motohashi, Alexei A Starobinsky, and Jun’ichi Yokoyama. f (r) gravity and its
cosmological implications. International Journal of Modern Physics D, 20(08):1347–1355, 2011.

– 22 –
[62] Yong-Seon Song, Wayne Hu, and Ignacy Sawicki. Large scale structure of f (r) gravity. Physical
Review D—Particles, Fields, Gravitation, and Cosmology, 75(4):044004, 2007.
[63] Kumar Ravi, Anirban Chatterjee, Biswajit Jana, and Abhijit Bandyopadhyay. Investigating
the accelerated expansion of the universe through updated constraints on viable f (r) models
within the metric formalism. Monthly Notices of the Royal Astronomical Society,
527(3):7626–7651, 2024.
[64] Suresh Kumar, Rafael C Nunes, Supriya Pan, and Priya Yadav. New late-time constraints on f
(r) gravity. Physics of the Dark Universe, 42:101281, 2023.
[65] Deng Wang. Pantheon+ constraints on dark energy and modified gravity: An evidence of
dynamical dark energy. Physical Review D, 106(6):063515, 2022.
[66] Kushal Lodha, Lucas Pinol, Savvas Nesseris, Arman Shafieloo, Wuhyun Sohn, and Matteo
Fasiello. Searching for local features in primordial power spectrum using genetic algorithms.
arXiv preprint arXiv:2308.04940, 2023.
[67] Ahana Kamerkar, Savvas Nesseris, and Lucas Pinol. Machine learning cosmic inflation.
Physical Review D, 108(4):043509, 2023.
[68] M Ballardini, Y Akrami, F Finelli, D Karagiannis, Baojiu Li, Y Li, Z Sakr, D Sapone,
A Achúcarro, M Baldi, et al. Euclid: The search for primordial features. Astronomy &
Astrophysics, 683:A220, 2024.
[69] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural networks,
61:85–117, 2015.
[70] Ziad Sakr and Matteo Martinelli. Cosmological constraints on sub-horizon scales modified
gravity theories with mgclass ii. Journal of Cosmology and Astroparticle Physics, 2022(05):030,
2022.
[71] SGOPAL Patro and Kishore Kumar Sahu. Normalization: A preprocessing stage. arXiv
preprint arXiv:1503.06462, 2015.
[72] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv preprint
arXiv:1803.08375, 2018.
[73] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.
Dropout: a simple way to prevent neural networks from overfitting. The journal of machine
learning research, 15(1):1929–1958, 2014.
[74] Santiago Casas, VF Cardone, Domenico Sapone, N Frusciante, Francesco Pace, Gabriele
Parimbelli, M Archidiacono, K Koyama, Isaac Tutusaus, Stefano Camera, et al. Euclid:
Constraints on f (r) cosmologies from the spectroscopic and photometric primary probes. arXiv
preprint arXiv:2306.11053, 2023.
[75] E Bellini, A Barreira, N Frusciante, B Hu, S Peirone, M Raveri, Miguel Zumalacarregui,
A Avilez-Lopez, M Ballardini, RA Battye, et al. Comparison of einstein-boltzmann solvers for
testing general relativity. Physical Review D, 97(2):023520, 2018.
[76] Yu Zhang, Peter Tiňo, Aleš Leonardis, and Ke Tang. A survey on neural network
interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence,
5(5):726–742, 2021.
[77] Scott M Lundberg, Gabriel Erion, Hugh Chen, Alex DeGrave, Jordan M Prutkin, Bala Nair,
Ronit Katz, Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee. From local explanations to
global understanding with explainable ai for trees. Nature machine intelligence, 2(1):56–67,
2020.
[78] Mukund Sundararajan and Amir Najmi. The many shapley values for model explanation. In
International conference on machine learning, pages 9269–9278. PMLR, 2020.

– 23 –
[79] Amirata Ghorbani and James Zou. Data shapley: Equitable valuation of data for machine
learning. In International conference on machine learning, pages 2242–2251. PMLR, 2019.
[80] Luis Fernando Machado Poletti Valle, Camille Avestruz, David J. Barnes, Arya Farahi,
Erwin T. Lau, and Daisuke Nagai. shaping the gas: understanding gas shapes in dark matter
haloes with interpretable machine learning. Mon. Not. Roy. Astron. Soc., 507(1):1468–1484,
2021.
[81] S. Gilda, S. Lower, and D. Narayanan. mirkwood: Fast and Accurate SED Modeling Using
Machine Learning. In American Astronomical Society Meeting Abstracts, volume 237 of
American Astronomical Society Meeting Abstracts, page 215.07D, January 2021.

– 24 –