WorksPhysics and Astronomy
Testing General Relativity by Reconstructing Dark Energy and the Growth Index with Physics-Informed Neural Networks
Abstract
We extend the Cosmo-PINN, a physics-informed neural network for model-independent cosmological reconstruction, to reconstruct the dark energy equation of state parameter and the growth index of matter perturbations. The reconstruction is performed through three modules: the background module, the matter growth module, and the growth index module. The background module learns the dark energy density from cosmological observations, and is reconstructed through the residual of the conservation equation. The reconstructed background is then passed to the other two modules, which reconstruct the observables and either directly from the linear perturbation equation of General Relativity or through an independently reconstructed growth index . Comparing the two reconstructions on the same background provides a consistency test of the underlying physical framework. The network is trained on the DESI DR2 Baryon Acoustic Oscillations, Cosmic Chronometers, Type Ia Supernovae, and redshift-space distortion data. Posterior uncertainties for the reconstructed cosmological parameters and functions are obtained through Hamiltonian Monte Carlo sampling. For the different data combinations, the posterior median of the present-day dark energy equation of state lies within the range , while the phantom divide crossing takes place at –. The posterior median of the present-day growth index reconstructed from the data lies within –, indicating suppression of the growth of cosmic structure.
Introduction
The cosmological constant , which leads to the standard CDM model, is the simplest candidate for the description of the cosmic acceleration [1–5]. It corresponds to a dark energy fluid with constant energy density and pressure, and consequently a constant equation of state parameter with the fixed value . Nevertheless, the cosmological constant suffers from the fine-tuning and coincidence problems [6], while the recent Baryon Acoustic Oscillation measurements from DESI DR2, combined with other late-time observational measurements, indicate a preference for a dynamical character of dark energy, in agreement with other independent analyses [9–11]. Furthermore, the cosmological constant faces observational tensions, the most important being the tension, where the value obtained from the Cosmic Microwave Background differs by from that obtained from local distance measurements [12–15]. These issues have motivated cosmologists to explore alternative theoretical approaches for the nature of dark energy and the origin of the cosmic acceleration, by introducing scalar fields and other matter sources [24, 25, 26, 27] [16, 17, 18, 19, 20, 21, 22, 23] as dark energy candidates, or modified theories of gravity [28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39], where the cosmic acceleration has a geometric origin driven by the geometric invariants which modify the Einstein-Hilbert action.
From a phenomenological approach, various parametric forms for the dark energy equation of state have been proposed in order to extract the information provided by cosmological observations. The most widely used is the Chevallier-Polarski-Linder (CPL) parametrization [40, 41], which can be seen as a first-order Taylor expansion of the dark energy equation of state around the present epoch. Various extensions and generalizations of the CPL model, which form the CDM family of models can be found in [42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58] and references therein.
A natural generalization of the CDM family of models is to expand the dark energy equation of state parameter in a suitable basis of functions, thereby allowing an agnostic reconstruction of its evolution [59, 60, 61, 62]. Such an approach was considered in [63], where the Padé approximation was applied for the reconstruction of the dark energy. The reconstruction was performed by constraining with the data the parametric forms of for different degrees of the expansion, and the different orders were compared through statistical criteria. More recently, in [64], the authors introduced an extension of the CPL parametrization by including higher-order terms of the Taylor expansion. It was found that the additional degrees of freedom are penalized by the Akaike information criterion, and that marginalizing over the additional terms beyond the CPL parametrization weakens the evidence for dynamical dark energy. A similar conclusion was obtained in [65] by using the Weighted Function Regression method.
Among the non-parametric approaches, Gaussian Processes (GP) have become one of the most popular tools for the reconstruction of the expansion history. In [66], the cosmographic parameters were reconstructed by using the GP method, while [67] presented a non-parametric reconstruction of the dark energy equation of state following the same approach. The dark energy-dark matter interaction was investigated using GP in [68]; for more applications of the GP in cosmology we refer the reader to [69, 70, 71, 72, 73, 74, 75, 76]. In GP, the reconstructed function is characterized by a mean and a covariance function. Nevertheless, the reconstruction is kernel dependent, and different selections for the kernel can lead to different reconstructed histories [90]. Moreover, reconstructed parameters such as the dark energy equation of state follow from higher-order derivatives of the reconstructed functions, which increases the uncertainties. Recently, machine learning techniques have found their way into applications in cosmology [77, 78, 79]. Genetic Algorithms [80, 81, 82] and Artificial Neural Networks (ANNs) [83, 84, 85, 86, 87, 88, 89] have been widely applied for the reconstruction of the Hubble function, the dark energy equation of state parameter and for the reconstruction of free functions in modified theories of gravity. Nevertheless, in the aforementioned approaches, the reconstruction is entirely data-driven, and the reconstructed functions are not required to satisfy the underlying physical laws governing the cosmological dynamics.
This limitation is addressed by Physics-Informed Neural Networks (PINNs) [91], where the physical laws are embedded in the loss function as residual terms [92]. Hence, during training, the neural network reconstructs the physical quantities from observational data while enforcing the underlying physical laws throughout the reconstruction domain. A PINN for the solution of the cosmological dynamics was recently presented in [93], while a generic gravitational consideration was introduced in [94]; see also [95, 96]. A PINN framework was introduced in [99] to reconstruct the supernova absolute magnitude. Parameter estimation for parametric dark energy models with the use of PINNs was recently described in [97]. As far as the equations for the linear cosmological perturbations are concerned, the PINN framework has been used before in [98] and in [100], where in the latter study PINNs were used for the joint reconstruction of the Hubble function and of the .
In [101] we introduced the Cosmo-PINN, a PINN framework for the reconstruction of the dark energy equation of state parameter from the background cosmological data. Within the Cosmo-PINN, the neural network learns the dark energy density , and the is reconstructed through the residual of the conservation equation. The cosmological parameters are trained jointly with the network, and the training is performed in two optimization stages, followed by a Bayesian stage in which the posterior distributions of the reconstructed functions and the cosmological parameters are calculated by using the Hamiltonian Monte Carlo approach. The first application of the Cosmo-PINN framework is presented in [102], where the varying index of the modified Barrow entropy was reconstructed from the late-time data.
Different theories of gravity can give rise to the same cosmic expansion history, making it difficult to distinguish between them using background cosmological observations alone. The growth of matter perturbations can be used to overcome this limitation, since it provides a complementary probe to the background expansion. Theories with the same background can predict different matter perturbations. The growth index defined as , has been introduced to quantify the difference among models with the same background evolution [103, 104]. For the CDM model, the growth index is approximately constant, , while values of close to the CDM value are also obtained for parametric dark energy models within General Relativity. However, has a strong dependence on the gravitational theory [105, 106, 107, 108, 109, 110], such that it has been used to test General Relativity on cosmological scales [111, 112, 113, 114, 115, 116, 117, 118, 119].
In this work, we extend the Cosmo-PINN to the growth of matter perturbations, and we introduce the redshift-space distortion (RSD) measurements as additional observational constraints. We introduce two independent modules: a growth module, in which the linear perturbation equation for the matter is introduced as a residual in the loss function and the network derives the and observables and the growth index ; in the second module, the growth index is reconstructed independently from the RSD data. The two modules share the same background; hence, the comparison of the and is a consistency test for the gravitational theory. We consider General Relativity, where the dark matter is described by a pressureless fluid source; thus, the new modules allow us to test General Relativity by using the RSD data. The structure of the paper is as follows.
In § 2 we present the cosmological framework, which is that of General Relativity in an isotropic and homogeneous FLRW background, with a dark energy fluid with a dynamical equation of state , which accounts for the late-time cosmic acceleration, and dark matter and baryons described by pressureless fluids. We introduce the linear matter perturbation equations and we define the growth index . The observational datasets used in this work are given in § 5. Furthermore, in § 6 we summarize the architecture of the PINN, together with the modifications that extend the Cosmo-PINN framework to the reconstruction of and . The growth index is reconstructed by two modules: from the perturbation equations of the matter, , for the underlying theory, and directly from the data of the matter perturbations. In Section V we present the reconstruction outcomes of the training process for different combinations of the datasets. Specifically, we investigate the effects of the supernova data on the cosmological parameters. The reconstructed and are presented with their credible intervals. Additionally, we derive the present-day values of the cosmological parameters and of the growth index . Finally, in § 9 we summarize our findings and discuss possible extensions of the Cosmo-PINN framework.
Dark Energy
Within the framework of General Relativity, we describe the universe by a FLRW background geometry
consistent with the cosmological principle, which assumes homogeneity and isotropy on sufficiently large scales. The function is the radius of the three-dimensional volume and is the lapse function. The comoving four-velocity is defined as , such that , from which we define the expansion rate by the expression , where , is the Hubble function. Without loss of generality, in the following we assume the gauge for the lapse function .
Let the spacetime be filled with a cosmic fluid described by the energy-momentum tensor
where , and denote the total energy density and pressure, respectively. The equation of state parameter for the total fluid is defined as .
The cosmic fluid consists of radiation, baryonic matter, and a dark sector comprising dark matter and dark energy. Radiation is described by the energy density and pressure , , the baryons by , , and the dark sector comprises the dark matter and the dark energy. Furthermore, the dark matter is considered to be given by a pressureless fluid source, , , and the dark-energy component is characterized by an energy density and the dynamical equation of state parameter , such that .
Hence, the energy-momentum tensor for the cosmic fluid is written equivalently
Background field equations
The Einstein field equations relate the gravitational field to the dynamics of matter. For the FLRW background, they lead to the Friedmann equations2
The Bianchi identity gives the condition3 , which provides the continuity equations for the cosmic fluid, that is,
In the scenario where the cosmic fluid components interact only gravitationally, from (6) it follows
Integrating and replacing in (4) we find the following expression for the Hubble function
where is the value for the Hubble function at the present , denotes the energy density of the fluid today, with the constraint
Growth of matter perturbations
In the following lines we discuss the basic elements of the evolution of the linear growth of matter perturbations and we introduce the growth index .
In our consideration, the total matter density consists of the dark matter and the baryons, that is, , and because there is no interaction term within the dark sector, the evolution of the matter density perturbations , is governed by the second-order linear differential equation [120, 121]
or equivalently,
where denotes the energy density of the total matter.
The growth rate of clustering is defined as . In terms of the new variables, equation (13) reads
where is the equation of state parameter for the cosmic fluid, that is,
In the late-time epoch of the universe, radiation can be neglected; thus, the latter equation reads
from which we observe that affects the clustering of the matter component through the background evolution. The first-order ordinary differential equation (17) is of Riccati type.
Consider now the growth index , defined as
then satisfies the equation
In this set of variables, for a slowly varying equation of state parameter , i.e. , it follows to first approximation [103, 104]
where for CDM, in which , we calculate , such that .
The growth index can be used as a phenomenological parameter to characterize deviations from the CDM, as well as from the case of a constant equation of state parameter. In the literature, various phenomenological parametrizations have been proposed to describe the evolution of the growth index. Some models that have been considered beyond the constant are, for instance, , , and others [111, 112, 113, 114, 115, 116, 117, 118, 119].
Cosmological Data
In our analysis, we combine background cosmological observations with the Redshift-Space Distortion (RSD) measurements from galaxy redshift surveys. From these data sets we can constrain both the background expansion history and the growth of matter perturbations.
The RSD data probe the growth of large-scale structure and constrain the evolution of matter perturbations. We make use of the catalogues presented in [122, 123], which provide measurements of the growth rate and of at given redshifts, where is the root-mean-square amplitude of matter fluctuations in spheres with radius , that is,
The measurements cover the redshift range , while the measurements span .
For the background constraints, we use Cosmic Chronometers (CC), Type Ia Supernovae (SNIa), and Baryon Acoustic Oscillations (BAO).
The CC catalogue provides measurements of the Hubble parameter at redshifts in the range , from old-aged galaxies [124, 125]. The likelihood is defined with the full covariance matrix4 as introduced in [126].
Moreover, we consider the BAO measurements as provided by the Dark Energy Spectroscopic Instrument Data Release 2 (DESI DR2) [7, 8]. The catalogue includes seven bins of measurements for the comoving angular distance, the volume-averaged distance and the Hubble distance normalized by the sound horizon at the baryon drag epoch , that is, , and .
Finally, we consider the PantheonPlus (PP) [127], Union3.0 (U3) [128] and DES-Dovekie (DD) [129] SNIa catalogues. The PP sample contains 1550 SNIa observations in the redshift range , while the U3 includes 2087 SNIa covering a comparable redshift range. DESD is the most recent catalogue, which follows from the reanalysis of the five-year Dark Energy Survey supernova program (DES-SN5YR), and includes 1820 SNIa events in the redshift range . We consider the PP sample with and without the Cepheid calibration from SH0ES.
Physics-Informed Neural Network
The architecture of the PINN adopted in this study is based on the Cosmo-PINN framework introduced in [101]. The Cosmo-PINN is designed to reconstruct the dark energy equation of state parameter from background cosmological observations, with the underlying physical laws imposed as constraints on the neural network rather than being solved independently by conventional numerical methods. This neural network architecture ensures that the reconstructed cosmological history remains consistent with the gravitational theory at every point of the redshift domain.
Within the Cosmo-PINN, the training provides a single optimal solution that minimizes the loss function, the maximum a posteriori (MAP) solution. Uncertainties of the reconstructed functions are derived in a Bayesian phase. For the derivation of the posterior, we perform sampling with the Hamiltonian Monte Carlo method, starting from the MAP solution. In this work, we extend the architecture of the Cosmo-PINN by introducing a dedicated growth module, such that the network provides the evolution of the matter density perturbations , which is constrained by the linear growth equation, and the growth rate and the combination , so that the RSD measurements can be used as data for the reconstruction process. Moreover, the extended network introduces an independent reconstruction of the growth index and of the growth observables through from the RSD data, without solving the master equation (13) for the perturbations; we shall refer to this as the growth index module.
The architecture of the PINN is schematically illustrated in Figure 1. The Cosmo-PINN consists of three stages: an Adam optimization stage, an L-BFGS optimization stage, and a Bayesian inference stage, which provides the posterior space for the reconstructed cosmological functions and cosmological parameters.
Figure 1. Architecture of the extended Cosmo-PINN. The original framework for the reconstruction of from background data is shown in blue. The growth module for and the growth-index module for introduced in this work are shown in orange.
Cosmo-PINN Background
Within the Cosmo-PINN framework, the redshift input is mapped to the unit interval via the transformation , in which , and is the largest value for the redshifts of the data sets, that is, . The multilayer perceptron (MLP) provides the function , which is related to the normalized dark energy density [101]
where the Hubble rate is given by the expression5
The definition of the with the exponential function is necessary in order to impose during the training.
The dark energy equation of state parameter is trained via the residual [101]
For the representation of the within the PINN, we employ two different approaches. We expand in a Chebyshev polynomial basis [130]
in which the coefficients are initialized at the CDM point, and for , so training begins from . In the second representation form, we allow to be described by an independent neural network, initialized at the cosmological constant limit. As discussed in [101], the two representations lead to statistically equivalent reconstructions of for a sufficiently large number of the Chebyshev expansion. The Chebyshev representation should not be confused with a parametric form of , but rather a series approximation of the function that follows from the trained dark energy density via the residual .
The SNIa compilations, the CC and the BAO catalogues are used to constrain the background, and the cosmological parameters , , are provided jointly with the Chebyshev coefficients for the dark energy equation of state parameter. The coefficients are initialized at the CDM point, and for , so training begins from . Hence, the network is trained to explore deviations from the CDM limit.
Additionally, the network provides a second output, the which is used in the training of the growth sector. The physical quantities and are passed to the growth sector as background quantities. Thus, all growth quantities are derived on the same background.
Growth module
For the growth module, the output from the main neural network is used to define through the formula
where now , and for close to the matter epoch . The parameter is related to the value of the growth rate at the matter epoch as . is a free parameter of the model, and it is trained by the RSD data; is the initial condition of the differential equation (15).
In order for the network to satisfy the physical law, the following residual term is introduced within the loss function
with a varying weight .
The theoretical observables and the are derived and constrained using the RSD data via the loss term
Furthermore, following the construction of the Cosmo-PINN, weak penalties on the first and second derivatives of the trained function are included in the loss function.
Furthermore, from the reconstructed , we calculate the growth index .
Growth index module
For the reconstruction of the growth index, we follow the same approach as for the reconstruction of the dark energy equation of state parameter, by using a Chebyshev expansion, or a network output.
We employ a Chebyshev expansion with the basis
where is initialized around the value. With this choice, we are able to track deviations from the CDM. Moreover, it is possible to express from a third neural network, similarly to the representation of , as presented in Figure 1.
Furthermore, we independently derive the growth rate and the corresponding observable directly from the reconstructed , where is given by the background, and now
Finally, the loss term is introduced
in order to drive the reconstruction from the data.
The RSD data constrain the function ; nevertheless, the growth index within General Relativity should satisfy the differential equation (19). We impose the physical law as a residual term
within the loss function with a constant weight , together with a soft penalty on the first derivative of the .
This module introduces two loss components: the from the data, and the of the residual. In the first two stages, the network is optimized to provide the function which minimizes the sum , and not necessarily the individual loss components, since the data do not necessarily satisfy the given physical law. If the RSD data do not support the given physical law, there is no function that can minimize both loss components. Nevertheless, in the Bayesian stage, only the is used to train the growth index . For the reconstructed background, we independently calculate the function from the differential equation (19). The deviation between the two independent reconstructed functions and can be seen as a test of whether the physical law (19) is supported by the data, that is, whether our cosmological framework of General Relativity with a pressureless description for the dark matter is in agreement with the given observations.
Training and Posterior Uncertainties
The network consists of three hidden layers with 64 neurons each, and as the activation function we consider . The training proceeds in two sequential phases: the Adam optimizer and the L-BFGS optimizer epochs. The Adam optimizer is employed for 15000 epochs and the L-BFGS optimizer is applied for up to 200 epochs, with early stopping once the total loss term does not decrease. The individual loss components differ by orders of magnitude in the size of their gradients [132]. Hence, in order to prevent a specific residual from driving the reconstruction, we employ the GradNorm method during the first 5000 epochs of the Adam optimizer to adjust the loss weights of the data sets and the residuals. After this epoch, the weights are frozen and remain fixed for the rest of the training.
Furthermore, when the solution has converged, the Bayesian stage takes place. The Hamiltonian Monte Carlo (HMC) [133] approach with the No-U-Turn Sampler (NUTS) [134] is considered, which allows us to sample the posterior distribution over the free cosmological parameters, namely , , , and , the Chebyshev coefficients of , the coefficients for the and the weights of the final network layer around the MAP solution, which provides the outputs and . For each posterior sample, the functions , , , are derived, while is calculated independently from equation (19). For the reconstructed functions and we consider the Chebyshev expansions of order and respectively. In the analysis presented in [101] it was found that there is no significant improvement in the reconstruction when we consider more terms in the expansion for the , in agreement with the result obtained in [136] for a polynomial basis for the reconstruction of the dark energy equation of state parameter. Moreover, for the we assume a smaller number, since the total number of RSD data points is relatively small.
The Gaussian priors for the free cosmological parameters and the Chebyshev coefficients are presented in Table I and Table II respectively. For the Chebyshev coefficients [131], we consider scalable priors, in order for the lower-order coefficients to have strong constraints and to avoid degeneracies as well as oscillatory behaviour; for more details, we refer to the discussion in [135].
| Parameter | Prior |
Table I. Gaussian priors for the cosmological parameters.
| Coefficient | Prior | Coefficient | Prior |
Table II. Gaussian priors for the Chebyshev coefficients. The priors are centered on the CDM and on the constant growth index .
Learning and
In this section, we present the results obtained from the training of the Cosmo-PINN on the reconstruction of the dark energy equation of state parameter and growth index . The training is performed for the following combinations of data sets: CC+BAO+RSD and CC+BAO+SNIa+RSD, where SNIa corresponds to the PP, U3, DD and PP+SH0ES catalogues.
CC+BAO+RSD
We train the Cosmo-PINN by using RSD data combined with the CC and BAO measurements. Figure 2 shows the evolution of the loss components during the first two stages of the training, that is, for the Adam-optimizer and L-BFGS optimizer epochs. We present the loss components for the background data, , and the RSD measurements , the residuals , which enforce the physical laws within the PINN, and the total loss function . The loss components decrease rapidly during the Adam stage, which means that the Cosmo-PINN fits the observational data with a solution that satisfies the physics residuals. Finally, the L-BFGS stage refines and further constrains the cosmological parameters , and as shown in Figure 3. Furthermore, in Figure 4 we plot the maximum absolute value of the PDE residual for the Adam-optimizer and L-BFGS optimizer epochs, from which we conclude that the reconstructed solution satisfies the physical law across the training period for the whole redshift range.
Figure 2. No SNIa: Evolution of the loss components during training for the combined data CC+BAO+RSD. The left panel corresponds to the data loss terms. The middle panel shows the physics residuals, and the right panel the total loss function. The vertical dotted line describes the transition to the L-BFGS optimizer stage.
Figure 3. No SNIa: Evolution of the cosmological parameters , , and of , during the Adam and L-BFGS optimizer training for the combined data CC+BAO+RSD. The vertical dotted line describes the transition to the L-BFGS optimizer stage.
Figure 4. No SNIa: Maximum absolute value of the PDE residual during training for the combined data CC+BAO+RSD. The vertical dotted line describes the transition to the L-BFGS optimizer stage.
In Figure 5 we present the reconstructed dark energy equation of state parameter with the 68% and 95% credible intervals (CI) as given by the Bayesian stage. For illustration, we present some individual trained solutions which were used to construct the posterior. The posterior median suggests a present value , while crosses the phantom divide line at . Moreover, Figure 6 summarizes the reconstructed evolution of the Hubble function , the deceleration parameter , and the dark energy density parameter . We also present the dark energy density , which is directly learned by the network and is used to derive the rest of the cosmological quantities.
Figure 5. No SNIa: Reconstructed dark energy equation of state parameter with the 68% and 95% CI and individual trained solutions. The reconstruction is for the combined data CC+BAO+RSD.
Figure 6. No SNIa: Reconstructed Hubble function , deceleration parameter , dark energy density and with the 68% and 95% CI and individual trained solutions. The reconstruction is for the combined data CC+BAO+RSD.
As far as the growth rate and are concerned, in Figure 7 we show the results of the reconstruction from the two modules, the growth module and the growth index module; the latter provides an individual reconstruction independent of the perturbation equation. We observe that for the two bands overlap, with the solution reconstructed by the module having tighter bands. However, at higher redshifts there are fewer data points, which do not allow for a proper reconstruction of the perturbation functions from the RSD data alone. This is not the case for the solutions obtained from the growth module, since the residual should always be satisfied, which drives the evolution of the perturbations. In Table III and Table IV we present the observable values of and as given by the data sets, and compare them with the reconstructed values through the two modules.
| via | via | ||
| 0.013 | |||
| 0.150 | |||
| 0.180 | |||
| 0.220 | |||
| 0.350 | |||
| 0.380 | |||
| 0.410 | |||
| 0.550 | |||
| 0.600 | |||
| 0.770 | |||
| 1.400 |
Table III. measurements from the RSD data and the reconstructed values from the growth and growth index modules. The reconstruction is for the combined data CC+BAO+RSD.
Figure 7. No SNIa: Reconstructed and from the and modules and the corresponding 68% and 95% CI. The reconstruction is for the CC+BAO+RSD combined data. The vertical dotted line indicates the maximum redshift at which RSD measurements are available.
| via | via | ||
| 0.020 | |||
| 0.025 | |||
| 0.067 | |||
| 0.100 | |||
| 0.150 | |||
| 0.320 | |||
| 0.380 | |||
| 0.440 | |||
| 0.570 | |||
| 0.590 | |||
| 0.700 | |||
| 0.730 | |||
| 0.740 | |||
| 0.760 | |||
| 0.850 | |||
| 0.978 | |||
| 1.050 | |||
| 1.400 | |||
| 1.480 | |||
| 1.944 |
Table IV. measurements from the RSD data and the reconstructed values from the growth and growth index modules. The reconstruction is for the combined data CC+BAO+RSD.
The growth index obtained from the two different reconstruction approaches is shown in Figure 8. We observe that the individual reconstruction of has very wide posterior bands in comparison to obtained from the module. This is because the growth index is not directly constrained by the data. The growth rate is constrained, which is related to through expression (18). Thus, at the matter epoch, , the growth rate becomes independent of and is driven to , as can easily be seen from the reconstruction in Figure 7. This limits the applicability of the growth index formalism at high redshifts. For small redshifts, , we observe that the lies outside the 68% CI of the ; nevertheless, it is always within the 95% interval, suggesting no tension between them.
Figure 8. No SNIa: We compare the individual reconstruction with , derived from the module, for the same reconstructed background. The reconstruction is for the CC+BAO+RSD combined data. The vertical dotted line indicates the maximum redshift at which RSD measurements are available.
Finally, in Figure 9 and Figure 10 we show the posterior distributions of the trained cosmological parameters , , , , , and of the Chebyshev coefficients and , respectively. We see that the coefficients have weak degeneracies, which follows from the selection of the shrinkage priors.
Figure 9. No SNIa: Marginalized distributions for the cosmological parameters , , , and , obtained from the HMC/NUTS analysis for the CC+BAO+RSD data. The contours correspond to the 68% and 95% credible regions.
Figure 10. No SNIa: Marginalized posterior distributions for the Chebyshev coefficients and entering the reconstruction of the dark energy equation of state parameter and of the growth index , obtained from the HMC/NUTS analysis for the CC+BAO+RSD data. The contours correspond to the 68% and 95% credible regions.
CC+BAO+SNIa+RSD
We extend the analysis by introducing the SNIa data in the training of the Cosmo-PINN. For the combined data set CC+BAO+SNIa+RSD, Table V summarizes the posterior medians of the cosmological parameters , , , and , and of the derived parameters and . It also shows the present-day values of the reconstructed equation of state , and of the growth index, , as obtained from both the and modules.
| Dataset | no SNIa | PP | U3 | DD | PP+SH0ES |
| H | |||||
Table V. Posterior median values and 68% credible intervals for the cosmological parameters obtained from the HMC analysis for five different dataset combinations.
The supernova data do not shift the median values of the cosmological parameters significantly, but mainly tighten the CI. The Hubble constant has median values within the range , with the largest value obtained when the SH0ES calibration is included. Furthermore, the energy density of the total matter today is derived within
the range . The parameters and , related to the amplitude of matter fluctuations and clustering, have posterior medians and . The trained parameter is always found to be lower than one, and the median values are within the range .
As far as the reconstructed functions and are concerned, the posterior medians of lie in the quintessence regime, , with the lowest value obtained when the SH0ES calibration is introduced. The growth index is tightly constrained when it is derived via the module, with and very small uncertainties for all combinations. This is because, in this approach, follows from the solution of the perturbation equation of General Relativity and thus it is fully determined by the reconstructed background. On the other hand, when is determined directly from the RSD data via the module, the obtained values lie in the range , with the CDM limit always within the 68% CI. The obtained values indicate a suppression of the growth of cosmic structure, in agreement with the analysis presented in [137]. However, its statistical significance is low, which means that the RSD data do not support a departure from General Relativity.
The results are visualized in Figure 11, from which we remark the consistency of the reconstruction of the cosmological parameters, independently of the supernova catalogue that is applied.
Figure 11. Comparison of the posterior median and the 68% CI for the cosmological parameters as provided by the training of the Cosmo-PINN and summarized in Table V. The left panel shows , and the right panel the present value of . The dots show the posterior median of as provided by General Relativity for the same background.
The reconstructed background function for the four combinations is presented in Figure 12, while in Figure 13 we show the reconstructed functions obtained by the two modules with their credible intervals.
Figure 12. CC+BAO+SNIa+RSD: Reconstructed dark energy equation of state parameter with the 68% and 95% CI and individual trained solutions. The reconstruction is for the combined data CC+BAO+SNIa+RSD.
Figure 13. CC+BAO+SNIa+RSD: Individual reconstruction of and for the same reconstructed background. The reconstruction is for the CC+BAO+SNIa+RSD combined data. The vertical dotted line indicates the maximum redshift at which RSD measurements are available.
For the different catalogues of SNIa, the today is within the quintessence regime, and it crosses the phantom divide line for all the datasets. Without SNIa data, the median takes the value at ; for the PP catalogue, the crossing occurs at , while the SH0ES calibration shifts the crossing to . Moreover, for the U3 catalogue the crosses the phantom divide line at , while the DD catalogue supports a later crossing at . Finally, for the growth index, the SNIa modify the dark energy equation of state parameter which affects the reconstructed ; however, the conclusions drawn before without SNIa data remain unchanged.
Conclusions
The Cosmo-PINN framework has been extended through the incorporation of modules that incorporate the RSD measurements into the data-driven reconstruction, allowing for the reconstruction of the growth index of matter perturbations. Specifically, the reconstructed background quantities, derived from and , are shared with two modules. In the matter growth module, the network learns the growth of matter , with the second-order linear perturbation equation of General Relativity introduced as a residual term in the
loss function. This module provides the observables and , which are constrained by the RSD data, while the General Relativity prediction for the growth index, , is also calculated. On the other hand, in the growth index module, the network learns directly from the RSD data, by using the already reconstructed background. This double reconstruction of the growth index allows us to perform a consistency test of General Relativity. The posterior distributions of the reconstructed functions and of the cosmological parameters were derived through HMC sampling around the MAP solution.
We trained the neural network by combining the RSD measurements with background data, specifically the CC, the DESI DR2 BAO, and three SNIa compilations, with the PP sample considered both with and without the SH0ES calibration. For all data set combinations employed in this work, the physics residuals reach small values across the entire redshift domain of the reconstruction, from which we conclude that the reconstructed solutions satisfy the differential equations that describe the physical laws. The outcome of the training is the dark energy density , which defines the Hubble function, the dark energy equation of state parameter , the growth of matter , and the growth indices and provided by the two modules. Furthermore, the posterior distributions of the cosmological parameters , , , and are derived from the HMC sampling. Although a Chebyshev expansion was considered for the representation of and ,
the reconstruction should be seen as non-parametric, since the main network learns , and the Chebyshev expansion represents the solution for that follows from the residual. Moreover, as demonstrated in [101], replacing the Chebyshev expansion of with a second neural network yields statistically equivalent results, with a slight increase in the uncertainties but a reduced computational cost.
From the analysis of the different data set combinations, we conclude that the introduction of the SNIa data slightly changes the median values of the background parameters, while it tightens their posteriors. In particular, we found , which increases to , when the SH0ES calibration is included. For the matter density, we find , and , such that the clustering parameter is , independently of the supernova catalogue. Moreover, the value of the dark energy equation of state parameter at present is found to be within the quintessence regime, , while the posterior median of crosses the phantom divide line at . As far as the two independent reconstructions of the growth index are concerned, the growth index derived from the matter growth module is tightly constrained, , since it is fully determined by the reconstructed background. In contrast, the growth index reconstructed directly from the RSD data has median values –, with significantly larger uncertainties. The two reconstructions are consistent within the 68% CI at the present epoch, while for , lies outside the 68% but always within the 95% credible interval of . The preference of the RSD data for is in agreement with results obtained previously [137]. Furthermore, the growth index is weakly constrained at high redshifts, since as independently of , which limits the applicability of the growth-index formalism during the matter-dominated era. Future measurements will significantly improve the constraints on , allowing for a more stringent test of General Relativity.
In future work, we plan to extend the Cosmo-PINN to modified theories of gravity and to interacting dark sector models.
Data Availability Statement: The Cosmo-PINN code is available upon reasonable request.
Acknowledgments
The author acknowledges support from FONDECYT Grant 1240514. AP acknowledges the COST Action CA23130 “Bridging high and low energies in search of quantum gravity (BridgeQG)”.
References
- [24]W. Yang, S. Zhang, D. Barbosa, M.A. Alvarez, S. Pan and L. Graef, Phys. Rev D 114, 063549 (2026)
- [25]M. Abedin, L.A. Escamilla, S. Pan, E. Di Valentino and W. Yang, Phys. Rev. D 112, 123537 (2025)
- [26]S. Sahlu, A. Paliathanasis, G. Leon and A. Abebe, Phys. Dark Univ. 52, 102280 (2026)
- [27]S. Sahlu, U. Mukhopadhyay, R.R. Mekuria and A. Abebe, MNRAS 546, staf1921 (2026)
- [28]H.A. Buchdahl, Mon. Not. Roy. Astron. Soc. 150, 1 (1970)
- [29]R. Ferraro and F. Fiorini, Phys. Rev. D 75, 084031 (2007)
- [30]L. Heisenberg, Physics Rept. 1066, 1 (2024)
- [31]T. Clifton, P.G. Ferreira, A. Padilla and C. Skordis, Phys. Rept. 513, 1 (2012)
- [32]S. Nojiri, S.D. Odintsov and M. Sasaki, Phys. Rev. D 71, 123509 (2005)
- [33]S. Sahlu, A. de la Cruz-Dombriz and A. Abebe, MNRAS 539, 690 (2025)
- [34]S. Nojiri, S.D. Odintsov and V.K. Oikonomou, Phys.Rept. 692, 1 (2017)
- [35]A. Paliathanasis, Phys. Dark Univ. 49, 101993 (2025)
- [36]A. Paliathanasis, Phys. Dark Univ. 52, 102337 (2026)
- [37]E.N. Saridakis, G.G. Gaetano, A. Lymperis and E.C. Telali, (2026) [arXiv:2609.28331]arxiv.org/abs/2609.28331
- [38]Y. Yang, Q. Wang, Z. Lu, D. Zhang, Y.-F. Cai and E.N. Saridakis, Sci. Bull. 59, 2698 (2024)
- [39]T.P. Sotiriou and V. Faraoni Rev. Mod. Phys. 82, 451 (2010)
- [40]M. Chevaller and D. Polarski, Int. J. Mod. Phys. D 10, 213 (2001)
- [41]E.V. Linder, Phys. Rev. Lett. 90, 091301 (2003)
- [42]L. Xing, J. Chen, Y. Gui, E. Schlegel and J. Lu, Mod. Phys. Lett. A 26, 885 (2011
- [43]M. Artola, I. Ayuso and R. Lazkoz, Phys. Rev. D 113, 023513 (2026)
- [44]W.J. Wolf, C. Garcia-Garcia and P.G. Ferreira, JCAP 05, 034 (2025)
- [45]A. Paliathanasis, Phys. Dark Univ. 48, 101956 (2025)
- [46]S. Lee, Annals of Physics 489, 170453 (2026)
- [47]Y. Carloni, O. Luongo and M. Biesiada, Phys. Rev. D 113, 103525 (2026)
- [48]Y. Carloni, O. Luongo and M. Biesiada, (2026) [arXiv:2608.01251]arxiv.org/abs/2608.01251
- [49]L.A. Escamilla, S. Pan, E. Di Valentino, A. Paliathanasis and J.A. Vazquez, Phys. Rev. D 111, 2 (2025)
- [50]S. Pan, W. Yang and A. Paliathanasis, Eur. Phys. J. C 80, 274 (2020)
- [51]N. Dimakis, A. Karagiorgos, A. Zampeli, A. Paliathanasis, T. Christodoulakis and Petros A. Terzis, Phys. Rev. D 93, 123518 (2016)
- [52]S. Nesseris and L. Perivolaropoulos, Phys. Rev. D 70, 043531 (2004)
- [53]J.-Z. Ma and X. Zhang, Phys. Lett. B 699, 233 (2011)
- [54]D.A. Kessler, L.A. Escamilla, S. Pan and E. Di Valentino, Phys. Rev. D 114, 063501 (2026)
- [55]W. Yang, E. Di Valentino, E.V. Linder, S. Zhang and S. Pan, Phys. Rev. D 114, 6 (2026)
- [56]S. Pan, E.N. Saridakis and W. Yang, Phys. Rev. D 98, 063510 (2018)
- [57]G. Montefalcone and R. Stiskalek, Phys. Rev. D 114, 043538 (2026)
- [58]M. Malekjani, S. Pourojaghi and Z. Davari, Phys. Rev. D 133, 103511 (2026)
- [59]P.K.S. Dunsby and O. Luongo, IJGMMP 13, 1630002 (2016)
- [60]D. Huterer and G. Starkman, Phys. Rev. Lett. 90, 031301 (2003)
- [61]R.G. Crittenden, G.-B. Zhao, L. Pogosian, L. Samushia and X. Zhang, JCAP 02, 048 (2012)
- [62]B.R. Dinda and R. Maartens, JCAP 01, 120 (2025)
- [63]A. Mehradi and S. Basilakos, EPJC 78, 889 (2018)
- [64]S. Nesseris, Y. Akrami and G.D. Starkman, Phys. Rev. D 114, L041305 (2026)
- [65]A. Gonzaalez-Fuentes and A. Gomez-Valent, JCAP 12, 049 (2025)
- [66]A. Shafieloo, A.G. Kim and E.V Linder, Phys. Rev. 85, 123530 (2012)
- [67]M. Seikel, C. Clarkson and M. Smith, JCAP 06, 036 (2012)
- [68]T. Yang, Z.-K. Guo and R.-G. Cai, Phys. Rev. D 91, 123533 (2015)
- [69]M.-J. Zhang and J.-Q. Xia, JCAP 12, 005 (2016)
- [70]J.F. Jesus, R. Valentim, A.A. Escobal and S.H. Pereira, JCAP 04, 053 (2020)
- [71]M.K. Yennapureddy and F. Melia, EPJC 78, 258 (2018)
- [72]R.C. Bernado and J.L. Said, JCAP 09, 014 (2021)
- [73]R. El Ouardi, A. Bouali, A. Errahmani, R.E. Keely, A. Shafieloo and T. Ouali, (2026) [arXiv:2604.04056]
- [74]M. Abedin, L.A. Escamilla, S. Pan, E. Di Valentino and W. Yang, Phys. Rev. D 112, 123537 (2025)
- [75]Y.-F. Cai, M. Khurshudyan and E.N. Saridakis, Astroph. J. 888, 62 (2020)
- [76]X. Ren, S.-F. Yan, Y. Zhao, Y.-F. Cai and E.N. Saridakis, Astrophys. J. 932, 131 (2022)
- [77]M. Ntampaka et al. (2019) [arXiv:1902.10159]arxiv.org/abs/1902.10159
- [78]L. Rojas, S. Espinoza, E. Gonzales, C. Maldonado and F. Luo, Galaxies 13, 114 (2025)
- [79]M. Lochner, J.D. McEwen, H.V. Peiris, O. Lahva and M.K. Winter, Astrophysical J. Supp. Ser. 225, 31 (2016)
- [80]R.C. Bernardo and Y. Chen, Astronomy and Computing 55, 101041 (2026)
- [81]R. Medel-Esquivel, I. Gomez-Vargas, A.A. Morales Sanchez, R. García-Salcedo and J. Alberto Vazquez, Universe 10, 11 (2024)
- [82]M. Peronaci, M. Martinelli and S. Nesseris, JCAP 08, 063 (2026)
- [83]G.-J. Wang, X.-J. Ma, S.-Y. Li and J.-Q. Xia, ApJS 246, 13 (2020)
- [84]J.-C. Zhang, Y. Hu, K. Jiao, H.-F. Wang, Y.-B. Xie, B. Yu, L.-L. Zhao and T.-J. Zhang, ApJS 270, 23 (2024)
- [85]I. Gomez-Vargas, R. Medel-Esquivel, R. Gacia-Salcedo and J.A. Vazquez, EPJC 83, 304 (2023)
- [86]K.F. Dialektopoulos, P. Mukherjee, J.L. Said and J. Mifsud, EPJC 83, 956 (2023)
- [87]K.F. Dialektopoulos, P. Mukherjee, J.L. Said and J. Mifsud, Physics Dark Universe 43, 101383 (2024)
- [88]P. Mukherjee, M.G. Dainotti, K.F. Dialektopoulos, J.L. Said and J. Mifsud, JHEAp 49, 100439 (2026)
- [89]Q. Zhang, G.-J. Wang and J.-Q. Xia, ApJ 1002, 128 (2026)
- [90]J.P. Johnson and H.K. Jassal, EPJC 85, 996 (2025)
- [91]M. Raissi, P. Perdikaris and G.E. Karniadakis, J. Comput. Phys. 378, 686 (2019)
- [92]G.R. Liu, PINN with Python: An Introduction, Scientech Publisher, Mason, USA (2025)
- [93]A.T. Chantada, S.J. Landau, P. Protopapas, C.G. Sóccola and C. Garrafo, Phys. Rev. D 107, 063523 (2023)
- [94]J.A. Carretero, J.F. Urban, F. Abalos, J. Masso, O.A. Reula, J.A. Pons and C. Palenzuela (2026) [arXiv:2609.38331]arxiv.org/abs/2609.38331
- [95]M. Yarahmadi and A. Salehi, JHEAp 50, 100498 (2026)
- [96]M. Yarahmadi and A. Salehi, EPJC 85, 1301 (2025)
- [97]A. Verna, S. Sourav, P.K. Aluri and D.F. Mota (2025) [arXiv:2508.12032]
- [98]L. Gomez, A.T. Chantada, S.J. Landau, C.G. Sóccola and P. Protopapas, Phys. Rev. D 112, 063515 (2025)
- [99]D. Staicova, Phys. Dark Univ. 52, 102342 (2026)
- [100]K. Dialektopoulos, (2026) [arXiv:2606.17614]arxiv.org/abs/2606.17614
- [101]A. Paliathanasis, (2026), [arXiv:2605.30139]arxiv.org/abs/2605.30139
- [102]S. Basilakos, A. Paliathanasis, E.N. Saridakis and S.A. Tsilioukas, (2026) [arXiv:2608.27469]arxiv.org/abs/2608.27469
- [103]L. Wang and P.J Steinhardt, Astrophys. J. 508, 483 (1998)
- [104]E.V. Linder, Phys. Rev. D 72, 043529 (2005)
- [105]F. Oliveira, B. Ribeiro, W.S. Hipolito-Ricaldi, F. Avila and A. Bernui, JCAP 12, 007 (2025)
- [106]M.A. Resco and A.L. Maroto, Phys. Rev. D 97, 043518
- [107]S. Basilakos and S. Nesseris, Phys. Rev. D 96, 063517 (2016)
- [108]S. Nesseris, G. Pantazis and L. Perivolaropoulos, Phys. Rev. D 96, 023542 (2017)
- [109]A.K. Sharma and M.M. Verma, Ap. J. 934, 13 (2022)
- [110]W. Khyllep, A. Paliathanasis and J. Dutta, Phys. Rev. D 103, 103521 (2021)
- [111]G. Ballesteros and A. Riotto, Phys. Lett. B. 668, 171 (2008)
- [112]P. Wu, H. Yu and X. Fu, JCAP 09, 019 (2009)
- [113]A. B. Belloso, J. Garcia-Bellido and D. Sapone, JCAP, 10, 10 (2011)
- [114]D. Polarski and R. Gannouji, Phys. Lett. B., 660, 439 (2008)
- [115]S. Basilakos and A. Pouri, MNRAS 423, 3761 (2012)
- [116]A. Pouri, S. Basilakos and M. Plionis, JCAP 08, 042 (2014)
- [117]G. Panotopoulos and A. Rincon, Phys. Rev. D 97, 103509 (2018)
- [118]S. Basilakos and S. Sola, Phys. Rev. D 92, 123501 (2015)
- [119]M. Aparicio Resco and A.L. Maroto, Phys. Rev. D 97, 043518 (2018)
- [120]P.J.E. Peebles, Principles of Physical Cosmology, Princeton University Press, New Jersey (1993)
- [121]L. Amendola and S. Tsujikawa, Dark Energy: Theory and Observations, Cambridge Univer- sity Press, Cambridge (2013)
- [122]F. Avila, A. Bernui, A. Bonilla, and R. C. Nunes, Eur. Phys. J. C 82, 594 (2022),
- [123]A. A. Escobal, H. A. P. Macedo, J. F. Jesus, R. C. Nunes, and J. A. S. Lima, Phys. Dark Univ. 52, 102320 (2026)
- [124]M. Moresco, R. Jimenez, L. Verde, A. Cimatti, and L. Pozzetti, Astrophys. J. 898, 82 (2020)
- [125]S. Vagnozzi, A. Loeb and M. Moresco, Astrophys. J. 908, 84 (2021)
- [126]M. Moresco, R. Jimenez, L. Verde, A. Cimatti, and L. Pozzetti, Astrophys. J. 898, 82 (2020)
- [127]D. Brout et al., The Astrophysical Journal 938, 110 (2022)
- [128]D. Rubin et al. (2023) [arXiv:2311.12098]arxiv.org/abs/2311.12098
- [129]B. Popovic et al. (DES) (2025), [arXiv:2511.07517] Biao Chen, Jing Wang, Hairun Xie, Qineng Wang, Shuai Zhang, Yifan Xia, Jifa Zhang (2026) [arXiv:2602.01737]DOI
- [131]R. Calderon et al. JCAP 10, 048 (2024)
- [132]Z. Chen, V. Badrinarayanan, C.-Y. Lee and A. Rabinovich, GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks, Proceedings of the 35th International Conference on Machine Learning 793 (2018)
- [133]R.M. Neal, Handbook of Markov Chain Monte Carlo, Chapman and Hall, CRC (2011)
- [134]M.D. Hoffman and A. Gelman, J. Machine Learning 15, 1593 (2014)
- [135]T. Park and G. Casella, J. American Stat. Association 103, 482 (2008)
- [136]A. Gonzalez-Fuentes and A. Gomez-Valent, JCAP 12, 049 (2025)
- [137]N.-M. Nguyen, D. Huterer and Y. Wen, Phys. Rev. Lett. 131, 111001 (2023)












