betaFeedback

WorksComputer Science

Improved Uncertainty Quantification in Physics-Informed Neural Networks Using Error Bounds and Solution Bundles

Abstract

Physics-Informed Neural Networks (PINNs) have been widely used to obtain solutions to various physical phenomena modeled as Differential Equations. As PINNs are not naturally equipped with mechanisms for Uncertainty Quantification, some work has been done to quantify the different uncertainties that arise when dealing with PINNs. In this paper, we use a two-step procedure to train Bayesian Neural Networks that provide uncertainties over the solutions to differential equation systems provided by PINNs. We use available error bounds over PINNs to formulate a heteroscedastic variance that improves the uncertainty estimation. Furthermore, we solve forward problems and utilize the obtained uncertainties when doing parameter estimation in inverse problems in cosmology.

Introduction

Physics-Informed Neural Networks (PINNs), first proposed by Lagaris et al. [1997], solve differential equations (DEs) by embedding the physics of the problem into the network, eliminating the need for extra data. PINNs offer advantages over traditional solvers: they are continuous, differentiable, and parallelizable, allowing them to bypass the need for previous time steps. However, challenges remain in computing solution errors or error bounds, with ongoing research [Liu et al., 2022, 2023, De Ryck and Mishra, 2022b]. Specific equations like Navier-Stokes and Elasticity are also under study [De Ryck and Mishra, 2022a, De Ryck et al., 2022, Guo and Haghighat, 2022].

Since their introduction, PINNs have rapidly gained attention, being applied to heat transfer [9], wave equations [Rasht-Behesht et al., 2022], and fluid mechanics [8] (Jin et al. 2021; Mao et al. 2020). Efforts to address failure modes and optimization issues are ongoing [19] (Nabian et al. 2021; Steger et al. 2022; Krishnapriyan et al. 2021), with interest in Bayesian PINNs for uncertainty quantification (UQ) [Yang et al., 2020, Linka et al., 2022, Graf et al., 2021, Psaros et al., 2023].

The use of NNs for solving ODEs and PDEs was pioneered by Lagaris et al. [1997] and later advanced by Raissi et al. [2019], who introduced PINNs for forward and inverse problems. Forward problems solve for the solution given the equation and boundary conditions, while inverse problems estimate unknown parameters. Inverse problems were implemented by Raissi et al. [2019] by treating these parameters as trainable variables.

Although effective, this approach lacks a robust mechanism for uncertainty quantification, which Bayesian Neural Networks (BNN) address by introducing distributions over network weights. However, simply applying Bayesian methods does not fully leverage the available information about solution accuracy. This is where Solution Bundles [Flamant et al., 2020] come into play. Solution Bundles enable statistical analysis over multiple possible solutions, providing a more comprehensive view of uncertainty in both the solution and the equation parameters.

In cosmology, DEs aim to explain the universe’s expansion. Testing new models against observations typically involves statistical analysis to determine parameter bounds. By combining Solution Bundles with BNNs, we enhance uncertainty quantification, offering a more reliable method for parameter estimation. Our contributions include:

  • Introducing error-bound-based heteroscedastic variance for better uncertainty quality.

  • Solving forward problems for cosmological equations while quantifying uncertainties.

  • Applying Solution Bundles to solve inverse problems for parameter estimation.

Background

Problem Formulation

We adopt a slightly different formulation from Psaros et al. [2022]. The DEs we will work with can be defined as follows:

Fλ[u(x)]=f(x),x∈Ω,(1)F_{\lambda}[u(\mathbf{x})] = f(\mathbf{x}), \quad\mathbf{x} \in\Omega, \tag*{(1)}
Bλ[u(x)]=b(x),x∈Γ,(2)B_{\lambda}[u(\mathbf{x})] = b(\mathbf{x}), \quad\mathbf{x} \in\Gamma, \tag*{(2)}

where x\mathbf{x} is the space-time coordinate, Ω\Omega is a bounded domain with boundary Γ\Gamma, f(x)f(\mathbf{x}) is the source term, uu is the solution of the system, FλF_{\lambda} is a differential operator, BλB_{\lambda} and b(x)b(\mathbf{x}) are the boundary conditions (BCs) operator and term, respectively, and λ\lambda denotes the parameters of the system.

In this paper, we will focus on problems where the operators FλF_{\lambda}, BλB_{\lambda} and terms f(x)f(\mathbf{x}), b(x)b(\mathbf{x}) are known. If λ\lambda is assumed to be known, the goal is to find the solution uu, referred to as the forward problem. Conversely, if uu is known and the aim is to estimate λ\lambda, then this is known as the inverse problem.

Physics-Informed Neural Networks

A Physics-Informed Neural Network uses a neural network, uθ(x)u_{\theta}(\mathbf{x}), to approximate the true solution u(x)u(\mathbf{x}) of a differential equations system. As discussed by Psaros et al. [2023], PINNs can be trained by minimizing a fitting dataset’s Mean Squared Error (MSE). A fitting dataset D={Df,Db}\mathcal{D} = \{\mathcal{D}_f, \mathcal{D}_b\} is composed of noisy observations of ff, Df={xi,fi}i=0Nf\mathcal{D}_f = \{\mathbf{x}_i, f_i\}_{i=0}^{N_f}, and noisy BCs data Db={xi,bi}i=0Nb\mathcal{D}_b = \{\mathbf{x}_i, b_i\}_{i=0}^{N_b}

L(θ)=wfNf∑i=0Nf∥fθ(xi)−fi∥22+wbNb∑i=0Nb∥bθ(xi)−bi∥22(3)\mathcal{L}(\theta) = \frac{w_f}{N_f}\sum_{i=0}^{N_f}\lVert f_{\theta}(\mathbf{x}_i)-f_i\rVert_2^2 + \frac{w_b}{N_b}\sum_{i=0}^{N_b}\lVert b_{\theta}(\mathbf{x}_i)-b_i\rVert_2^2 \tag*{(3)}

Where wfw_f and wbw_b are weighting constants. Psaros et al. [2023] call this setup the Forward Deterministic PDE Problem, they also describe Mixed Deterministic PDE Problem where the objective is to obtain solutions for uu and λ\lambda, and the Mixed Stochastic PDE Problem which deals with stochastic PDEs.

In this work, we adopt a different optimization problem to solve the forward deterministic problem and address the inverse problem separately rather than solving both forward and inverse problems simultaneously.

Solving Forward Problems

We begin by defining the residual of a differential equation as

rθ(x)=Fλ[uθ(x)]−f(x)(4)r_{\theta}(\mathbf{x}) = F_{\lambda}[u_{\theta}(\mathbf{x})] - f(\mathbf{x}) \tag*{(4)}

The computation of Fλ[uθ(x)]F_{\lambda}[u_{\theta}(\mathbf{x})] is easy to implement thanks to automatic differentiation provided by Deep Learning frameworks such as PyTorch [Paszke et al., 2019]. This formulation allows PINNs to be trained as a self-supervised network. Since any solution u∗u^{*} to the differential equation satisfies Fλ[u∗(x)]−f(x)=0F_{\lambda}[u^{*}(\mathbf{x})] - f(\mathbf{x}) = 0, we train the network to minimize the square of the residual

min⁡θ1Nr∑iNrrθ2(xi).(5)\min_{\theta}\frac{1}{N_r}\sum_i^{N_r}r_{\theta}^2(\mathbf{x}_i). \tag*{(5)}

Usually, xi\mathbf{x}_i are sampled from Ω\Omega with a uniform distribution or by taking an equally-spaced subset. It is important to note that we do not deal with noisy data, unlike the work by Raissi et al. [2019].

Enforcing Boundary Conditions

While adding a term for BCs in the loss (see Eq. (3)) when dealing with data is a good way to incorporate such knowledge, there is no guarantee that the conditions will be satisfied. When BCs are known rather than observed in the data, we can use the transformation introduced by Lagaris et al. [1997] that enforces the BCs to be always satisfied. This is achieved by writing the approximate solution as a sum of two terms:

u~θ(x):=A(x)+F(x,uθ(x))(6)\tilde{u}_{\theta}(\mathbf{x}) := A(\mathbf{x}) + F(\mathbf{x}, u_{\theta}(\mathbf{x})) \tag*{(6)}

where AA does not depend on the network parameters θ\theta and it satisfies the BCs. Since we need u~θ\tilde{u}_{\theta} to satisfy BCs, FF is constructed so it does not contribute to them. This transformation is also used by Graf et al. [2021], Chen et al. [ [12].

One-dimensional Initial Value Problem Given an initial condition u0=u(t0)u_0 = u(t_0), we consider a transformation

u~θ(t):=u0+(1−e−(t−t0))uθ(t)(7)\tilde{u}_{\theta}(t) := u_0 + (1-e^{-(t-t_0)})u_{\theta}(t) \tag*{(7)}

In Appendix A of the Supplementary Material we show the enforcing of two-dimensional Dirichlet BCs. Similar transformations can be defined for Neumann and mixed BCs.

Solution Bundles

Solution Bundles [Flamant et al., 2020] extend PINNs by allowing the network to take equation parameters λ∈Λ\lambda\in\Lambda as inputs. This modification allows the network to approximate a variety of solutions to a parameterized differential equation without the need to retrain for each value of λ\lambda.

In § 2.1 we considered a unique value for λ\lambda and one BC term b(x)b(\mathbf{x}). When working with Solution Bundles instead, we have subsets for the equation parameters and BCs.

λ∈Λ⊂Rp\lambda\in\Lambda\subset\mathbb{R}^{p}
b(x)∈B(x)⊂Rnb(\mathbf{x}) \in B(\mathbf{x}) \subset\mathbb{R}^{n}

Λ\Lambda, B(x)\mathbb{B}(\mathbf{x}) are such subsets that will be used to train the NN. pp and nn are the dimensionality of the equation parameters and the system’s state variable, respectively. The transformation described in § 2.4 can also be used for Solution Bundles. Eq. (6) turns into:

u~θ(x,λ):=A(x,λ)+F(x,uθ(x,λ))(8)\tilde{u}_{\theta}(\mathbf{x},\lambda) := A(\mathbf{x},\lambda) + F(\mathbf{x},u_{\theta}(\mathbf{x},\lambda)) \tag*{(8)}

Training Solution Bundles

Flamant et al. [2020] proposed a weighting function for the residual loss (Eq. (5)) to influence how the approximation error is distributed across the training region. However, in this work, for simplicity, we stick to the unweighted residual loss

r~θ:=Fλ[uθ(x,λ)]−f(x)(9)\tilde{r}_{\theta} := \mathcal{F}_{\lambda}[u_{\theta}(\mathbf{x},\lambda)] - f(\mathbf{x}) \tag*{(9)}
min⁡θ∑iN∑jMr~θ2(xi,λj).(10)\min_{\theta}\sum_{i}^{N}\sum_{j}^{M}\tilde{r}_{\theta}^{2}(\mathbf{x}_{i},\lambda_{j}). \tag*{(10)}

where we have redefined rθr_{\theta} from Eq. (4) for the Solution Bundle case as r~θ\tilde{r}_{\theta}.

Solution Bundles for Solving Inverse Problems

So far we have explained how to solve forward problems, that is, finding a solution to a differential equations system. As we described, for the case of PINNs, the solving step is an optimization problem.

On the other hand, inverse problems aim to find a differential system that best describes some collected data. For this, we assume the differential system can explain the phenomena observed and we seek to estimate the system’s parameters for a given dataset.

An effective approach to addressing the parameter estimation problem involves statistical analysis, specifically through the application of Bayesian methods. This approach necessitates multiple computations of the system’s solution, corresponding to each parameter value. Traditional numerical methods require the discretization and integration process to be performed for each of these solutions. In contrast, Solution Bundles eliminate the need for retraining for each parameter value, thereby expediting the computational process. A comprehensive explanation of the probabilistic setup for Bayesian parameter estimation is provided in § 5.

Error Bounds for PINNs

Good quality uncertainties should correlate with the true error of a solution. Since the true error is not accessible, we use error bounds in § 4.2 to improve the uncertainty in the Bayesian NNs.

In [Liu et al., 2022, 2023], the authors present algorithms for computing error bounds on PINNs. These bounds apply to linear ODEs, systems of linear ODEs, non-linear ODEs in the form ϵvk,1\epsilon v^{k,1}, as well as certain types of PDEs. These algorithms are independent of the NN architecture and depend solely on the structure of the equation as defined in Eq. (1) and the residuals of the DE.

For the Solution Bundle setup, the network error is denoted as η(x,λ):=u(x,λ)−u~θ(x,λ)\eta(\mathbf{x},\lambda) := u(\mathbf{x},\lambda) - \tilde{u}_{\theta}(\mathbf{x},\lambda), and the error bound is represented by a scalar function B\mathbb{B} such that

∥η(x,λ)∥≤B(x,λ)(11)\lVert\eta(\mathbf{x},\lambda)\rVert\leq\mathbb{B}(\mathbf{x},\lambda) \tag*{(11)}

In this work we use the error bounds developed for first-order linear ODEs with nonconstant coefficient. In Appendix B of the Supplementary Material provide the algorithm for its computation and for first-order linear ODEs with constant coefficient. Appendix B.3 and Algorithm 1 describe how obtain tight bounds.

Bayesian Neural Networks

To quantify uncertainty, we adopt a Bayesian perspective on the neural networks. We do this by viewing the neural network as a probabilistic model p(y∣x,θ)p(y\mid\mathbf{x},\theta) and placing a prior distribution over its parameters p(θ)p(\theta). Using Bayes’ theorem, we can get the posterior distribution of θ\theta:

p(θ∣D)=p(D∣θ)p(θ)p(D)(12)p(\theta\mid\mathcal{D}) = \frac{p(\mathcal{D}\mid\theta)p(\theta)}{p(\mathcal{D})} \tag*{(12)}

p(D∣θ)p(\mathcal{D}\mid\theta) is the likelihood distribution over a dataset D={(xi,yi)}i=1N\mathcal{D} = \{(\mathbf{x}_{i},y_{i})\}_{i=1}^{N}.

The posterior distribution allows us to make predictions about unseen data by taking expectations. Consider a new data point x^\hat{\mathbf{x}}, we can obtain the probability of the output yy being y^\hat{y} as:

p(y^∣x^,D)=Ep(θ∣D)[p(y^∣x^,θ)]=∫Θp(y^∣x^,θ)p(θ∣D)dθ(13)p(\hat{y}\mid\hat{\mathbf{x}},\mathcal{D}) = \mathbb{E}_{p(\theta\mid\mathcal{D})}[p(\hat{y}\mid\hat{\mathbf{x}},\theta)] = \int_{\Theta}p(\hat{y}\mid\hat{\mathbf{x}},\theta)p(\theta\mid\mathcal{D})d\theta \tag*{(13)}

Usually, the integral in Eq. (13) is analytically intractable, and we have to resort to Monte Carlo (MC) approximations that can be computed as:

p(y^∣x^,D)≈∑i=1Mp(y^∣x^,θi), where θi∼p(θ∣D)(14)p(\hat{y}\mid\hat{\mathbf{x}},\mathcal{D}) \approx\sum_{i=1}^{M}p(\hat{y}\mid\hat{\mathbf{x}},\theta_{i}), \text{ where }\theta_{i}\sim p(\theta\mid\mathcal{D}) \tag*{(14)}

Shortcomings of Residual-Based UQ Methods in PINNs

A direct application of Bayesian Neural Networks to PINNs is straightforward by placing the likelihood over the residuals rr of the PINN solution given a coordinate point x\mathbf{x} i.e. p(r∣x,θ)p(r\mid\mathbf{x},\theta). For more details see § 4.3.

However, evaluating a PINN solution based only on the residual loss Eq. (5) can be misleading. A low residual at x\mathbf{x} does not guarantee a low solution error at x\mathbf{x}. In Control Theory, for instance, an Integral Controller drives steady-state error to zero, yet errors may still exist:

u(t)=ki∫0te(τ) dτ(15)u(t) = k_i \int_0^t e(\tau)\,d\tau \tag*{(15)}

where uu is the control input, e(τ):=r(τ)−y(τ)e(\tau) := r(\tau)-y(\tau) the error, rr the reference, yy the system output, and kik_i the integral gain [Åström and Murray, 2008]. For a DE dudt=f(t)\frac{du}{dt}=f(t) with IC u(t0)=u0u(t_0)=u_0, the approximation error u^(τ)−u(τ)\hat{u}(\tau)-u(\tau) can be expressed as:

u^(τ)−u(τ)=∫t0τR(t) dt(16)\hat{u}(\tau)-u(\tau)=\int_{t_0}^{\tau}R(t)\,dt \tag*{(16)}

where R(t):=du^dt−f(t)R(t):=\frac{d\hat{u}}{dt}-f(t), given the approximation satisfies the ICs. Proof of Eq. (16) is in Proposition F.1 of the Supplementary Material. Eq. (16) suggests that the residual at some point x\mathbf{x} is not enough to characterize the solution error at the same point x\mathbf{x}.

These challenges, along with other Failure Modes of PINNs, remain an active area of research [Krishnapriyan et al., 2021, Wang et al., 2022, Penwarden et al., 2023]. Notably, Wang et al. [2022] identify an implicit bias in the PINN framework that can severely violate the temporal causality of dynamical systems. As a result, residual minimization at a given time tit_i may occur even when predictions at earlier times are inaccurate.

In Figure 1 we see how a PINN solution has near maximum and minimum error for the same residual value. This illustrates the decoupling of residuals from solution errors, which can be attributed to a violation of temporal causality. Furthermore, Figure 13 to Figure 16 show how there is no clear correlation between residuals and solution errors. These figures were obtained by training a deterministic PINN on a cosmological model, the construction details are provided in Appendix § 18.

Normalized Absolute Residual vs Normalized Absolute Relative Error of a Deterministic NN Solution for CPL Model

Figure 1. Normalized Absolute Residual vs Normalized Absolute Relative Error of a Deterministic NN Solution for CPL Model.

Relationship between solution errors and residuals from $\Lambda$CDM

Figure 13. Relationship Between Solution Errors and Residuals from Λ\LambdaCDM. Samples were collected as described in I.

Relationship between solution errors and residuals from HS

Figure 16. Relationship Between Solution Errors and Residuals from HS. Samples were collected as described in I. Colors differentiate variables in the system.

Our experiments demonstrate that the baseline method produces uncalibrated predictive distributions and, in some cases, fails to approximate the true solutions. To address these limitations, we propose a two-step approach that incorporates error bounds into the solutions, leading to predictive distributions with improved calibration.

Two-Step Bayesian PINNs

We use a two-step approach to obtain uncertainties in the solutions of equations. In the first step, we train a PINN

as a Solution Bundle that we refer to as the deterministic network. We denote this network as uθdet⁡:Ω,Λ→Ru_{\theta_{\det}}:\Omega,\Lambda\to\mathbb{R} and θdet⁡∗\theta_{\det}^{*} as the parameters resulting after training.

Note that as shown in Eq. (10), this training step is carried out without the use of any data, but rather sampling from the domain (Ω×Λ)(\Omega\times\Lambda) and minimizing the network’s residuals. Figure 2 shows a diagram of the entire training process.

Diagram of Both Training Steps

Figure 2. Diagram of Both Training Steps. In the Second Step θdet∗\theta^{*}_{\mathrm{det}} are The Resulting Network Parameters From the First Step and L\mathcal{L} is The Likelihood Function.

Bayesian Neural Network Training

In the second step, we use the outputs of uθdet⁡∗u_{\theta_{\det}^{*}} as targets to train a Bayesian Neural Network.

We construct a dataset by taking the space-time coordinates and equation parameters as independent variables, and the outputs of the deterministic net as dependent variables:

D={(xi,λi,uθdet⁡∗(xi,λi))∣xi∈Ω,λi∈Λ,uθdet⁡∗(xi,λi)∈R}i=1N′\mathcal{D}=\left\{(\mathbf{x}_i,\lambda_i,u_{\theta_{\det}^{*}}(\mathbf{x}_i,\lambda_i))\mid\mathbf{x}_i\in\Omega,\lambda_i\in\Lambda,u_{\theta_{\det}^{*}}(\mathbf{x}_i,\lambda_i)\in\mathbb{R}\right\}_{i=1}^{N'}

We define a BNN uθu_\theta by assigning distributions to the network parameters and dataset as it is done in [Graf et al., 2021]. For all three methods described in § 2.7 we use a Gaussian prior:

θ∼N(0,σprior)\theta\sim\mathcal{N}(0,\sigma_{\mathrm{prior}})

To formulate the likelihood, we assume a Gaussian additive noise on the observations2

uθdet⁡∗(x,λ)=uθ(x,λ)+η(x,λ),(17)u_{\theta_{\det}^{*}}(\mathbf{x},\lambda)=u_\theta(\mathbf{x},\lambda)+\eta(\mathbf{x},\lambda), \tag*{(17)}

where η(x,λ)∼N(0,σlike(x,λ))\eta(\mathbf{x},\lambda)\sim\mathcal{N}(0,\sigma_{\mathrm{like}}(\mathbf{x},\lambda))

This setup is often used in machine learning, but we employ it with a different interpretation; see § 4.2. In this paper, we model the standard deviation σLike(x,λ)\sigma_{\mathrm{Like}}(x,\lambda) as known, but it can also be modeled as unknown. Given Eq. (17), the resulting likelihood is:

p(D∣θ)=∏i=1N′N(uθ(x,λ),σLike(x,λ))(18)p(\mathcal{D}|\theta) = \prod_{i=1}^{N'} \mathcal{N}(u_{\theta}(x,\lambda),\sigma_{\mathrm{Like}}(x,\lambda)) \tag*{(18)}

We assume Gaussian noise in Eq. (17) due to its simplicity, tractability (see § 4.1.1), and suitability as a starting point for evaluating our methodology. To assess this assumption, we analyze the distributional behavior of uθdetu_{\theta_{\mathrm{det}}}, with details provided in Appendix § 18. Figure 4 to Figure 12 illustrate the solution distributions obtained by a PINN for various cosmological models. Although these distributions deviate from a perfect Gaussian, they are not significantly distant, suggesting that a Gaussian approximation is a reasonable initial choice.

Distribution of $x_m(z)$ from $\Lambda$CDM

Figure 4. Distribution of xm(z)x_m(z) from Λ\LambdaCDM. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Distribution of $r(z)$ from HS across iterations and values of $t$

Figure 12. Distribution of r(z)r(z) from HS. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Posterior Distribution Approximation Methods

To apply the MC approximation in Eq. (13), we must generate samples from the posterior distribution p(θ∣D)p(\theta|\mathcal{D}). Given the complexity of neural networks and the need for computational efficiency, we compare three methods: Neural Linear Models (NLMs), Bayes By Backpropagation (BBB), and Hamiltonian Monte Carlo (HMC). NLMs offer a lightweight approach by approximating the posterior in a linearized feature space, making them suitable for speed-critical applications. BBB uses variational inference to achieve greater accuracy while maintaining tractability for larger networks. HMC is the most advanced method, producing highly accurate samples at the expense of increased computational complexity. Comparing these techniques allows for an evaluation of the trade-offs between accuracy and computational feasibility.

Neural Linear Models (NLMs) NLM represent Bayesian linear regression using a neural network for the feature basis, with only the final layer parameters treated as stochastic. Using Gaussian prior and likelihood allows for tractable inference, leading to the posterior and predictive distributions given by:

p(u∣x,λ,D)=N(μNLM(x,λ),σNLM(x,λ))(19)p(u \mid x,\lambda,\mathcal{D}) = \mathcal{N}(\mu_{\mathrm{NLM}}(x,\lambda),\sigma_{\mathrm{NLM}}(x,\lambda)) \tag*{(19)}
μNLM(x,λ)=Φθ(x,λ)μpost\mu_{\mathrm{NLM}}(x,\lambda) = \Phi_{\theta}(x,\lambda)\mu_{\mathrm{post}}
σNLM(x,λ)=σLike2(x,λ)+Φθ(x,λ)ΣpostΦθT(x,λ)\sigma_{\mathrm{NLM}}(x,\lambda) = \sigma_{\mathrm{Like}}^{2}(x,\lambda) + \Phi_{\theta}(x,\lambda)\Sigma_{\mathrm{post}}\Phi_{\theta}^{T}(x,\lambda)

where μpost\mu_{\mathrm{post}} and Σpost\Sigma_{\mathrm{post}} are posterior parameters, and Φθ(x,λ)\Phi_{\theta}(x,\lambda) is the learned feature map. The details of μpost\mu_{\mathrm{post}} and Σpost\Sigma_{\mathrm{post}} computation can be found in Appendix § 12 of the Supplementary Material.

Bayes By Backpropagation (BBB) BBB approximates the posterior of network parameters using variational inference (VI) and minimizes the KL divergence between the true posterior and a variational distribution q(θ∣ρ)q(\theta|\rho). This results in an approximation q(θ∣ρ)q(\theta|\rho) that is used to generate Monte Carlo samples for p(θ∣D)p(\theta|\mathcal{D}). We assume independent parameters under a mean-field approximation [4], defining q(θ∣ρ)=∏i=1Nq(θi∣ρi)q(\theta|\rho) = \prod_{i=1}^{N} q(\theta_i|\rho_i), using a Gaussian distribution for the variational posterior [5].

Hamiltonian Monte Carlo (HMC) HMC employs a Metropolis-Hastings algorithm [Metropolis et al., 2004, Hastings, 1970] to draw samples from p(θ∣D)p(\theta|\mathcal{D}) by simulating particle movement through Hamiltonian dynamics. To enhance sampling efficiency, we utilize the No-U-Turn Sampler [Hoffman and Gelman].

Improving Predictive Uncertainty with Error Bounds

When learning from observations y^\hat{y}, the standard approach assumes these observations have an error ϵ\epsilon:

y^=yw(x)+ϵ(20)\hat{y} = y_w(x) + \epsilon \tag*{(20)}

where yw(x)y_w(x) is a deterministic function parameterized by ww. This formulation attributes all errors to the observations. However, we can introduce an additional error term η\eta for model error:

y^=yw(x)+η+ϵ(21)\hat{y} = y_w(x) + \eta+ \epsilon \tag*{(21)}

Separating model error η\eta from observational error ϵ\epsilon is typically challenging due to their additive nature. If we know there is no observational error, we can use Eq. (21) similarly to Eq. (20). This is the case in Section 4.1, where the dataset has observations uθ∗det⁡(x,λ)u_{\theta^*}^{\det}(x,\lambda) free of noise:

Examples of ΛCDM Bayesian solutions obtained using the Bundle Solver

Figure 3. Examples of ΛCDM Bayesian Solutions Obtained Using the Bundle Solver. Analytic Solutions are Presented in Dotted Lines.

uθ∗det⁡(x,λ)=uθ(x,λ)+ϵ+η(22)u_{\theta^*}^{\det}(x,\lambda) = u_\theta(x,\lambda) + \epsilon+ \eta \tag*{(22)}

However, the approximate solution uθ∗det⁡u_{\theta^*}^{\det} may still have prediction errors. We model this error with a Gaussian distribution with standard deviation σLike(x,λ)\sigma_{\mathrm{Like}}(x,\lambda), which we treat as known in two ways:

  1. Homoscedastic:

σLike(x,λ)={0x∈Γ,const⁡x∉Γ\sigma_{\mathrm{Like}}(x,\lambda) = \begin{cases} 0 & x \in\Gamma, \\ \operatorname{const} & x \notin\Gamma \end{cases}

We set σLike\sigma_{\mathrm{Like}} to be zero on the boundaries because we enforce BC always to be met.

  1. Error Bounds Based Heteroscedastic:

σLike(x,λ)=B(x,λ)\sigma_{\mathrm{Like}}(x,\lambda) = \mathbb{B}(x,\lambda)

This choice ensures we are taking into account the error made by uθ∗det⁡u_{\theta^*}^{\det} in the predictive uncertainty.

Baseline

In the literature, the standard approach to formulating the forward problem with Bayesian PINNs relies on noisy observations of u(x)u(x), f(x)f(x) and b(x)b(x) [Psaros et al., 2022, Zou et al., 2023, Yang et al., 2020]. However, since our setup does not incorporate observations, we adopt as a baseline a Bayesian neural network with the likelihood defined over the residuals:

p(Dr∣θ)=∏i=1NrN(rθ(x,λ),σr)(23)p(\mathcal{D}_r \mid\theta) = \prod_{i=1}^{N^r} \mathcal{N}(r_\theta(x,\lambda),\sigma_r) \tag*{(23)}

The training procedure follows the approach described in Section 4.1, with the exception that we could not implement NLM in this setting. This limitation arises because the residuals depend on the solution, thereby breaking the analytical tractability required for NLM.

Probabilistic Formulation of Inverse Problems

The Bayesian framework can be used to estimate the parameters λ\lambda of a DE system. For a given set of observations O={(xi,μi,σi)∣xi∈Ω,μi∈R,σi∈R+}i=1O\mathcal{O} = \{(x_i,\mu_i,\sigma_i) \mid x_i \in\Omega,\mu_i \in\mathbb{R},\sigma_i \in\mathbb{R}^{+}\}_{i=1}^{O}, we seek to find the posterior distribution p(λ∣O)p(\lambda\mid\mathcal{O}) of the parameters. Here μi\mu_i and σi\sigma_i are the mean and standard deviation of the observations at some point xix_i, respectively.

Assume we have a probability distribution over the solutions p(uλ(x)∣x,λ)p(u_\lambda(x) \mid x,\lambda), then using Bayes’ Theorem, the posterior can be computed as p(λ∣O)=p(O∣λ)/p(λ)p(\lambda\mid O)=p(O\mid\lambda)/p(\lambda). The likelihood p(O∣λ)p(O\mid\lambda) is obtained by marginalizing over the solutions as:

p(xi,μi,σi∣λ)=∫up(xi,μi,σi∣uλ(xi))⋅p(uλ(xi)∣xi,λ) du(24)p(x_i,\mu_i,\sigma_i\mid\lambda)=\int_u p(x_i,\mu_i,\sigma_i\mid u_\lambda(x_i))\cdot p(u_\lambda(x_i)\mid x_i,\lambda)\,\mathrm{d}u \tag*{(24)}
≈1M∑j=1Mp(μi,σi∣uλ(j)(xi))(25)\approx\frac{1}{M}\sum_{j=1}^{M}p(\mu_i,\sigma_i\mid u_\lambda^{(j)}(x_i)) \tag*{(25)}
p(O∣λ)=∏i=1Op(xi,μi,σi∣λ)(26)p(O\mid\lambda)=\prod_{i=1}^{O}p(x_i,\mu_i,\sigma_i\mid\lambda) \tag*{(26)}

where uλ(j)∼p(uλ(xi)∣xi,λ)u_\lambda^{(j)}\sim p(u_\lambda(x_i)\mid x_i,\lambda).

The distribution induced by FCNNs can be interpreted as a Delta distribution, where the function learned by the network outputs a specific value uθdet(x)u_{\theta_{\mathrm{det}}}(x) for a given input xx. In contrast, BNNs naturally provide a distribution over solutions through their posterior predictive, thus Eq. (24) results in a marginalization over the posterior predictive distribution p(u(xi,λ)∣xi,λ,D)p(u(x_i,\lambda)\mid x_i,\lambda,\mathcal{D}).

We use a uniform distribution to aim for an uninformative prior for λ\lambda. In the same way Chantada et al. [ [11] did, to define the likelihood p(O∣λ)p(O\mid\lambda), we assume the observations are normally distributed around the true solution, i.e., μi∼N(u(xi),σi)\mu_i\sim\mathcal{N}(u(x_i),\sigma_i).

Having defined the prior and likelihood, we can apply a sampling algorithm to approximate p(λ∣O)p(\lambda\mid O). We use the emcee Python package [Foreman-Mackey et al., 2013] which implements the samplers introduced by Goodman and Weare [2010].

Experiments on Cosmology Models

This section specifies the cosmological equations we used to test our methodology. These equations were solved with PINNs by Chantada et al. [ [11]. For brevity, we describe the Λ\LambdaCDM and Parametric Dark Energy models. The details of Quintessence and f(R)f(R) gravity (or HS for the name of the authors Hu and Sawicki) are provided in Appendix § 16. However, Table 1 lists the variables and parameters of each cosmological model.

EquationTypeVariablesParameters
ΛCDMLinear Equationxmx_mΩm,0∗,H0\Omega^{*}_{m,0}, H_0
CPLLinear EquationxDEx_{DE}w0∗,w1∗,Ωm,0,H0w_0^{*}, w_1^{*}, \Omega_{m,0}, H_0
QuintessenceNon-linear Systemx,yx,yλ∗,Ωm,0∗,H0\lambda^{*}, \Omega^{*}_{m,0}, H_0
HSNon-linear Systemx,y,v,Ω,rx,y,v,\Omega,rb∗,Ωm,0∗,H0b^{*}, \Omega^{*}_{m,0}, H_0

Table 1. Cosmology Equations Variables and Parameters. Parameters Marked With * are Inputs in the Bundle Solution, The Remainder are Needed to Compute the Hubble Parameter.

Λ\LambdaCDM The equation and initial conditions are

dxmdz=3xm1+z\frac{\mathrm{d}x_m}{\mathrm{d}z}=\frac{3x_m}{1+z}
xm(z=0)=Ωm,0x_m(z=0)=\Omega_{m,0}

where Ωm,0\Omega_{m,0} is a parameter. Having a solution to this equation, the Hubble parameter can be obtained as

H(z)=H0xm(z)+1−Ωm,0(27)H(z)=H_0\sqrt{x_m(z)+1-\Omega_{m,0}} \tag*{(27)}

here H0H_0 is also a parameter.

Parametric Dark Energy We refer to this model as CPL for the names of the authors Chevallier, Polarski and Linder [13] (Linder 2003). The equation and initial conditions are

dxDEdz=3xDE1+z(1+w0+w1z1+z)\frac{\mathrm{d}x_{DE}}{\mathrm{d}z}=\frac{3x_{DE}}{1+z}\left(1+w_0+\frac{w_1z}{1+z}\right)
xDE(z=0)=1−Ωm,0x_{DE}(z=0)=1-\Omega_{m,0}

where w0w_0, w1w_1 and Ωm,0\Omega_{m,0} are parameters. Having a solution to this equation, the Hubble parameter can be obtained as

H(z)=H0Ωm,0(1+z)3+xDE(z)(28)H(z)=H_0\sqrt{\Omega_{m,0}(1+z)^3+x_{DE}(z)} \tag*{(28)}

here H0H_0 is also a parameter.

For brevity, we provide the error bounds computation details in Appendix § 11 of the Supplementary Material.

Forward Problems We trained uθdet∗u_{\theta^{*}_{\mathrm{det}}} using methods from [11], then built a dataset D\mathcal{D} to train the BNNs. We compared homoscedastic and error-bounds-based heteroscedastic variance for Λ\LambdaCDM and CPL equations, while providing homoscedastic variance results for Quintessence and HS. The networks were trained using regular PINNs and Solution Bundles, referred to as Forward and Bundle, respectively, with implementation details in Appendix § 17 of the Supplementary Material. We provide the complete implementation in a code repository.3

We evaluated the NN solutions by calculating the Median Relative Error (MRE) against analytical solutions for Λ\LambdaCDM and CPL, and numerical solutions for Quintessence and HS, using a Runge-Kutta method. We also computed the median residual, and miscalibration area (MA) [14] to assess uncertainty quality. Results are summarized in Table 2, with additional metrics in Table 13 to Table 15.

EquationMethodMedian REMedian ResidualMiscal. Area
Λ\LambdaCDMFCNN0.0010.0-
BBB0.9780.1470.491
HMC0.0840.1980.4
NLM + 2S0.0188.1570.183
BBB + 2S0.0350.9630.143
HMC + 2S0.0040.1970.123
NLM + 2S + EB0.00210.7830.063
BBB + 2S + EB0.0472.4760.05
HMC + 2S + EB0.0030.2720.098
CPLFCNN0.00.001-
BBB0.1280.0140.199
HMC0.00.010.317
BBB + 2S0.0630.0060.15
HMC + 2S0.0040.0050.255
BBB + 2S + EB0.0330.0030.177
HMC + 2S + EB0.0110.0040.145
Quint.FCNN0.0070.0-
BBB0.0940.0120.151
HMC0.0020.0050.147
NLM + 2S0.0960.1750.136
BBB + 2S0.120.0110.119
HMC + 2S0.0160.0020.048
HSFCNN0.0010.0-
BBB0.3930.0830.449
HMC0.2260.010.455
NLM + 2S0.2591.6660.315
BBB + 2S0.2960.130.396
HMC + 2S0.2860.3450.486

Table 2. Evaluation Metrics of All Bundle Networks and Equations in the Testing Region. Here we Use 2S for Two-step and EB for Error Bounds.

EquationMethodForward Q10Q20Q30Q40Q50Q60Q70Q80Q90Q100Bundle Q10Q20Q30Q40Q50Q60Q70Q80Q90Q100
ΛCDMFCNN0.00.00.00.00.00.1620.3940.5580.6690.7460.00.00.00.0010.0010.0170.2480.480.6470.819
ΛCDMNLM0.00.00.0010.0010.0070.1840.4180.5770.6850.7590.0010.0020.0030.0060.0180.0580.320.5540.7010.843
ΛCDMBBB0.00.00.00.0010.0040.1620.3970.560.6710.7480.00.00.0010.0010.0020.0170.2360.470.6390.814
ΛCDMHMC0.0010.0010.0020.0040.0190.2280.4430.5930.6960.7670.0010.0010.0020.0030.0070.0510.2120.4150.5970.799
ΛCDMNLM + EB0.00.00.00.00.00.170.4060.5690.6790.7540.00.00.0010.0010.0020.0170.1840.3810.5370.717
ΛCDMBBB + EB0.00.00.00.00.00.1560.3910.5590.6660.750.0010.0020.0040.0050.0070.0250.2540.490.6570.835
ΛCDMHMC + EB0.00.00.00.00.0040.2020.4320.580.690.7650.00.00.0010.0010.0020.0330.1520.3530.5410.78
CPLFCNN0.00.00.00.00.00.00.0010.0030.0060.0130.00.00.00.0010.0010.0170.2480.480.6470.819
CPLBBB0.00.00.00.00.00.00.0010.0030.0060.0110.00.0010.0020.0060.0730.3020.6842.28398.7389.53e+08
CPLHMC0.00.00.00.00.0010.0080.0220.0410.0660.0950.0010.0020.0040.0070.0330.130.3631.711100.0038.31e+08
CPLBBB + EB0.00.00.00.00.00.00.00.0010.0040.0090.0020.0040.0070.0140.0340.2390.7432.33321.779.01e+08
CPLHMC + EB0.00.00.00.00.00.010.0380.0780.1240.1730.0020.0040.0060.010.0210.3050.99928.5676.67e+053.88e+26
Quint.FCNN0.00.00.00.00.0010.0060.0390.1030.1710.2230.0010.0010.0020.0040.0070.0120.0210.0390.0770.26
Quint.NLM0.0030.0060.0090.0160.0340.0580.0940.160.2370.2880.0140.0270.0420.0650.0960.1450.2110.3250.708164.163
Quint.BBB0.0010.0020.0030.0050.0120.0270.0670.1290.1970.9180.0010.0030.0050.0080.0130.020.0340.0570.10424.699
Quint.HMC0.0120.020.0260.0430.0650.1350.2180.290.3481.310.0010.0030.0050.0080.0150.0260.0440.0710.11918.676
HSFCNN0.00.00.00.00.00.1620.3940.5580.6690.7460.00.00.00.0010.0010.0170.2480.480.6470.819
HSNLM0.0030.0070.010.0150.0260.9058.20323.65740.486147.1480.0780.1350.1820.2130.2470.6353.40316.61988.0251.33e+09
HSBBB0.0020.0030.0030.0050.0080.0220.0310.0420.1190.8310.0390.0890.1430.230.4290.7151.4343.4639.2117.52e+05
HSHMC0.1210.3050.5210.6740.7421.3082.717.06616.96359.1250.0670.1230.1690.2170.2940.3680.6292.2918.9429.06e+05

Table 13. Error Quantiles Of The NN Solutions Relative To The Analytical Solution For ΛCDM And CPL, And The Numerical Solution For Quintessence And HS.

EquationMethodMREMean ResidualMiscal. AreaRMS Cal.MA Cal.SharpnessNLLCRPSCheckIntervalAcc. MAEAcc. RMSEAcc. MDAEAcc. MARPD
Λ\LambdaCDMFCNN0.1780.0------------
Λ\LambdaCDMBBB0.940.1990.4910.5660.4860.07667151.12724.69212.347257.60124.7336.87314.224181.346
Λ\LambdaCDMHMC0.1965.1690.40.4550.3960.822139.80610.3095.161102.63410.6122.1420.75225.336
Λ\LambdaCDMNLM + 2S0.2068.2780.1830.2110.1820.45604.06812.2156.111124.94312.36925.2050.03229.855
Λ\LambdaCDMBBB + 2S0.2045.3860.1430.1730.1410.487935.00611.8295.918121.11911.96524.3620.09228.573
Λ\LambdaCDMHMC + 2S0.1715.4460.1230.1480.1221.439.0610.3465.182100.2210.80922.8730.01123.62
Λ\LambdaCDMNLM + 2S + EB0.1459.7140.0630.0740.06325.004−0.5325.742.89826.9439.35720.0030.01519.212
Λ\LambdaCDMBBB + 2S + EB0.1985.170.050.0580.0525.0081.6477.0043.53631.72211.42123.2840.29426.712
Λ\LambdaCDMHMC + 2S + EB0.1685.1040.0980.1220.09725.033−0.2436.363.21228.81410.59722.3230.02923.026
CPLFCNN0.0620.001------------
CPLBBB0.6831410.6270.20.2220.1985777.129−2.2613545.8961777.48633236.5553770.00776032.1270.00528.948
CPLHMC0.3621722.4660.3170.3410.3141.05e+05−6.2921324.918668.9597053.0561969.83641625.8110.013.055
CPLBBB + 2S2.751818.4870.150.1640.1492121.188.8855054.0622529.0251204.5875148.73596419.6810.00437.724
CPLHMC + 2S15.5332213.8850.2550.2720.2522.92e+05−5.2112891.451459.77718215.132540.45245826.6170.031.464
CPLBBB + 2S + EB2.09e+059.58e+080.1770.2120.1752.85e+0926693.4182.68e+081.35e+082.10e+093.09e+081.23e+100.00361.241
CPLHMC + 2S + EB2.35e+057.49e+050.1450.1650.1432.89e+0690.9492.74e+051.38e+052.06e+063.20e+051.03e+070.00147.367
QuintessenceFCNN0.0240.027------------
QuintessenceBBB0.1280.0790.1480.1930.1470.0047.61e+090.0170.0090.1590.0190.0410.002nan
QuintessenceHMC0.0090.0190.1470.1650.1460.0077.61e+090.0020.0010.0090.0020.0070.0nan
QuintessenceNLM + 2S0.390.330.1360.1580.1350.00651.9050.0390.0190.3750.040.0990.00224.652
QuintessenceBBB + 2S0.1590.120.1190.1650.1180.0047.61e+090.0270.0140.2650.0290.0640.002nan
QuintessenceHMC + 2S0.0410.0370.0480.0540.0470.0148.50e+090.0050.0030.0260.0070.020.0nan
HSFCNN0.1780.0------------
HSBBB0.6630.1360.4490.5180.4450.0155.90e+120.2640.1322.6490.270.3430.227nan
HSHMC0.4080.1520.4550.520.450.0415.90e+120.1260.0641.0030.1450.1930.125nan
HSNLM + 2S3740.0572.27e+060.3150.3630.3120.02720707.1050.380.193.9260.3840.5510.26441.024
HSBBB + 2S0.4170.3610.3960.4540.3920.0495.90e+120.20.1011.8240.2160.3490.113nan
HSHMC + 2S2.7393.4560.4860.5610.4820.0035.98e+120.1360.0681.4010.1370.2010.096nan

Table 15. Metrics of the NN Bundle Solutions.

For the CPL model, we utilized a reparameterization from [11] that separates equation parameters from Bundle Network training, although this method is incompatible with Neural Linear Models (NLM).

Inverse Problems For the inverse problem, we used 30 measurements of the Hubble parameter HH from the Cosmic Chronometers (CC) method [16] (Simon et al. 2005; Stern et al. 2010; Moresco et al. 2012; Moresco 2015; Moresco et al. 2016). Each measurement includes a tuple (zi,Hobs,σHobs)(z_i, H^{obs}, \sigma_{H^{obs}}), indicating redshift, observed mean Hubble parameters, and their standard deviation. The CC dataset is available in Table 9 of the Supplementary Material, with Eq. (41) used as the likelihood of observations.

zzH(z)±σHH(z) \pm\sigma_Hkm/s/Mpc\mathrm{km/s}/\mathrm{Mpc}Ref.
0.0969±1269 \pm12
0.1783±883 \pm8
0.2777±1477 \pm14
0.495±1795 \pm17
0.9117±23117 \pm23[Simon et al., 2005]
1.3168±17168 \pm17
1.43177±18177 \pm18
1.53140±14140 \pm14
1.75202±40202 \pm40
0.4897±6297 \pm62[Stern et al., 2010]
0.8890±4090 \pm40
0.179175±475 \pm4
0.199375±575 \pm5
0.351983±1483 \pm14
0.5929104±13104 \pm13[Moresco et al., 2012]
0.679792±892 \pm8
0.7812105±12105 \pm12
0.8754125±17125 \pm17
1.037154±20154 \pm20
0.0769±19.669 \pm19.6
0.1268.6±26.268.6 \pm26.2
0.272.9±29.672.9 \pm29.6
0.2888.8±36.688.8 \pm36.6[16]
1.363160±33.6160 \pm33.6
1.965186.5±50.4186.5 \pm50.4[Moresco, 2015]
0.380283±13.583 \pm13.5
0.400477±10.277 \pm10.2
0.424787.1±11.287.1 \pm11.2[Moresco et al., 2016]
0.449792.8±12.992.8 \pm12.9
0.478380.9±980.9 \pm9

Table 9. Measurements Of The Hubble Parameter H Using The Cosmic Chronometers Technique

We estimated equation parameters using the Solution Bundles from the forward step and performed inference with the emcee package [Foreman-Mackey et al., 2013], running 32 chains for 10,000 steps, resulting in 320,000 samples per parameter. Table 11 and Table 12 show the results and their concordance with values found in the literature, respectively.

EquationMethodw0w_0w1w_1λ\lambdabbΩm,0\Omega_{m,0}H0H_0
Λ\LambdaCDMFCNN----0.31±0.050.31 \pm0.0568.5±2.6368.5 \pm2.63
Λ\LambdaCDMBBB----0.1±0.010.1 \pm0.0179.95±0.7579.95 \pm0.75
Λ\LambdaCDMHMC----0.34±0.010.34 \pm0.0168.23±0.468.23 \pm0.4
Λ\LambdaCDMNLM + 2S----0.33±0.010.33 \pm0.0167.5±0.5167.5 \pm0.51
Λ\LambdaCDMBBB + 2S----0.33±0.010.33 \pm0.0167.42±0.6867.42 \pm0.68
Λ\LambdaCDMHMC + 2S----0.32±0.010.32 \pm0.0168.05±0.468.05 \pm0.4
Λ\LambdaCDMNLM + 2S + EB----0.32±0.010.32 \pm0.0168.14±0.3668.14 \pm0.36
Λ\LambdaCDMBBB + 2S + EB----0.31±0.010.31 \pm0.0168.42±0.4368.42 \pm0.43
Λ\LambdaCDMHMC + 2S + EB----0.32±0.010.32 \pm0.0168.16±0.3768.16 \pm0.37
CPLFCNN−1.03±0.3-1.03 \pm0.3−2.49±2.55-2.49 \pm2.55--0.34±0.120.34 \pm0.1265.76±7.6165.76 \pm7.61
CPLBBB−1.0±0.25-1.0 \pm0.25−2.53±3.06-2.53 \pm3.06--0.32±0.130.32 \pm0.1366.08±6.4766.08 \pm6.47
CPLHMC−1.05±0.3-1.05 \pm0.3−2.46±2.88-2.46 \pm2.88--0.32±0.140.32 \pm0.1465.8±7.0665.8 \pm7.06
CPLNLM + 2S−1.06±0.32-1.06 \pm0.32−2.91±2.76-2.91 \pm2.76--0.3±0.160.3 \pm0.1664.33±7.5864.33 \pm7.58
CPLBBB + 2S−1.01±0.3-1.01 \pm0.3−2.82±2.48-2.82 \pm2.48--0.31±0.160.31 \pm0.1665.01±6.6365.01 \pm6.63
CPLHMC + 2S−1.07±0.33-1.07 \pm0.33−3.02±2.45-3.02 \pm2.45--0.34±0.150.34 \pm0.1566.74±7.2766.74 \pm7.27
CPLNLM + 2S + EB−0.91±0.29-0.91 \pm0.29−2.61±2.63-2.61 \pm2.63--0.44±0.140.44 \pm0.1463.53±6.4663.53 \pm6.46
CPLBBB + 2S + EB−1.06±0.25-1.06 \pm0.25−2.24±2.85-2.24 \pm2.85--0.32±0.110.32 \pm0.1166.41±6.766.41 \pm6.7
CPLHMC + 2S + EB−1.12±0.28-1.12 \pm0.28−2.57±3.18-2.57 \pm3.18--0.32±0.140.32 \pm0.1467.65±6.6267.65 \pm6.62
Quint.FCNN--1.24±1.01.24 \pm1.0-0.27±0.070.27 \pm0.0767.36±7.1467.36 \pm7.14
Quint.BBB--0.88±1.160.88 \pm1.16-0.29±0.070.29 \pm0.0766.33±5.2866.33 \pm5.28
Quint.HMC--0.92±1.150.92 \pm1.15-0.3±0.060.3 \pm0.0667.96±4.4367.96 \pm4.43
Quint.NLM + 2S--1.36±0.831.36 \pm0.83-0.32±0.10.32 \pm0.164.5±4.9564.5 \pm4.95
Quint.BBB + 2S--1.13±1.191.13 \pm1.19-0.3±0.050.3 \pm0.0568.05±5.4668.05 \pm5.46
Quint.HMC + 2S--0.8±1.140.8 \pm1.14-0.3±0.060.3 \pm0.0665.7±5.6265.7 \pm5.62
HSFCNN---2.53±1.872.53 \pm1.870.28±0.070.28 \pm0.0769.03±7.9469.03 \pm7.94
HSBBB---1.82±1.951.82 \pm1.950.3±0.060.3 \pm0.0666.75±6.7566.75 \pm6.75
HSHMC---2.06±1.962.06 \pm1.960.27±0.080.27 \pm0.0866.73±6.566.73 \pm6.5
HSNLM + 2S---2.34±1.842.34 \pm1.840.28±0.120.28 \pm0.1257.44±8.8157.44 \pm8.81
HSBBB + 2S---1.98±1.871.98 \pm1.870.28±0.070.28 \pm0.0772.78±10.0172.78 \pm10.01
HSHMC + 2S---2.29±1.672.29 \pm1.670.28±0.090.28 \pm0.0967.82±10.3867.82 \pm10.38

Table 11. Parameters Mean And Standard Deviation Estimation Of All Cosmology Models.

EquationMethodw0w_0w1w_1λ\lambdabbΩm,0\Omega_{m,0}H0H_0
ΛCDMFCNN----1σ1\sigma1σ1\sigma
ΛCDMBBB----19σ19\sigma15σ15\sigma
ΛCDMHMC----5σ5\sigma3σ3\sigma
ΛCDMNLM----4σ4\sigma4σ4\sigma
ΛCDMBBB----3σ3\sigma3σ3\sigma
ΛCDMHMC----4σ4\sigma3σ3\sigma
ΛCDMNLM + EB----4σ4\sigma3σ3\sigma
ΛCDMBBB + EB----3σ3\sigma2σ2\sigma
ΛCDMHMC + EB----4σ4\sigma3σ3\sigma
CPLFCNN2σ2\sigma1σ1\sigma--1σ1\sigma1σ1\sigma
CPLBBB3σ3\sigma1σ1\sigma--1σ1\sigma2σ2\sigma
CPLHMC2σ2\sigma1σ1\sigma--1σ1\sigma2σ2\sigma
CPLNLM2σ2\sigma1σ1\sigma--1σ1\sigma2σ2\sigma
CPLBBB2σ2\sigma2σ2\sigma--1σ1\sigma2σ2\sigma
CPLHMC2σ2\sigma2σ2\sigma--1σ1\sigma1σ1\sigma
CPLNLM + EB3σ3\sigma1σ1\sigma--2σ2\sigma2σ2\sigma
CPLBBB + EB2σ2\sigma1σ1\sigma--1σ1\sigma1σ1\sigma
CPLHMC + EB2σ2\sigma1σ1\sigma--1σ1\sigma1σ1\sigma
Quint.FCNN--2σ2\sigma-1σ1\sigma-
Quint.BBB--1σ1\sigma-1σ1\sigma-
Quint.HMC--1σ1\sigma-1σ1\sigma-
Quint.NLM--2σ2\sigma-1σ1\sigma-
Quint.BBB--1σ1\sigma-1σ1\sigma-
Quint.HMC--1σ1\sigma-1σ1\sigma-
HSFCNN---2σ2\sigma1σ1\sigma1σ1\sigma
HSBBB---1σ1\sigma1σ1\sigma1σ1\sigma
HSHMC---1σ1\sigma1σ1\sigma1σ1\sigma
HSNLM---2σ2\sigma1σ1\sigma2σ2\sigma
HSBBB---1σ1\sigma1σ1\sigma1σ1\sigma
HSHMC---2σ2\sigma1σ1\sigma1σ1\sigma

Table 12. Smallest Number Of Sigmas Within Which The Results From [18, 1] (Motta et al. 2021) Fall.

Discussion

The FCNN consistently achieves the lowest median RE and mean residual across all equations, outperforming Bayesian methods in terms of accuracy. This highlights a tradeoff between equipping PINNs with uncertainty quantification and maintaining accuracy. In this work, our primary focus is on uncertainty rather than pure accuracy, as a well-calibrated model with higher error is generally more desirable than an uncalibrated model with low error. Properly calibrated uncertainties enable a meaningful assessment of prediction reliability.

From Table 2, we observe that the baseline Bayesian methods—BBB and HMC with a residual likelihood—exhibit significantly higher errors compared to the deterministic network, particularly for the CPL and HS equations. This effect is especially pronounced for BBB, which fails to approximate the Λ\LambdaCDM solution. Additionally, both methods exhibit high miscalibration areas, indicating that their uncertainty estimates are poorly calibrated.

Introducing our two-step (2S) method substantially improves calibration, as evidenced by the reduction in miscalibration areas across all cases. For instance, applying 2S to BBB in the Λ\LambdaCDM equation reduces the miscalibration area from 0.491 to 0.143.

Further incorporating error bounds into the two-step method enhances uncertainty calibration even further. Notably, BBB + 2S + EB achieves the lowest miscalibration area (0.05) for the Λ\LambdaCDM equation, while HMC + 2S + EB provides the best calibration in CPL (0.145). However, certain cases exhibit excessively large mean residuals, such as in CPL, where BBB + 2S + EB results in a residual of 9.58×1089.58 \times10^{8}. This suggests that while error bounds improve calibration, they may introduce numerical instability in the metrics due to some samples’ extremely large errors.

For the Quintessence equation, HMC + 2S achieves the best calibration (0.048) while maintaining a low median RE (0.016), demonstrating an effective balance between accuracy and uncertainty quantification. In contrast, for the HS equation, while the two-step method improves calibration, overall performance remains suboptimal, with high median RE and residuals persisting across all Bayesian approaches.

In Table 3 and Table 4, we present results separately for the training and OOD regions. In the training region, the two-step approach consistently improves both accuracy and miscalibration, and the addition of error bounds further enhances calibration across most equations. For example, in the Λ\LambdaCDM case, combining two steps with error bounds yields the lowest miscalibration area among all methods, indicating effective uncertainty correction when the model operates within its training distribution.

EquationMethodMedian REMedian ResidualMiscal. Area
ΛCDMFCNN0.00.0-
ΛCDMBBB0.9610.120.491
ΛCDMHMC0.0410.070.358
ΛCDMNLM + 2S0.0034.2120.402
ΛCDMBBB + 2S0.0080.2380.269
ΛCDMHMC + 2S0.0010.0380.307
ΛCDMNLM + 2S + EB0.0014.2110.016
ΛCDMBBB + 2S + EB0.010.2280.248
ΛCDMHMC + 2S + EB0.0010.0140.251
CPLFCNN0.00.001-
CPLBBB0.0520.0160.161
CPLHMC0.00.0090.482
CPLBBB + 2S0.0240.0090.085
CPLHMC + 2S0.0010.0080.436
CPLBBB + 2S + EB0.0070.0050.14
CPLHMC + 2S + EB0.0010.0080.159
Quint.FCNN0.0030.0-
Quint.BBB0.0590.0050.09
Quint.HMC0.0010.0030.279
Quint.NLM + 2S0.0630.0970.326
Quint.BBB + 2S0.0470.0020.165
Quint.HMC + 2S0.0060.00.179
HSFCNN0.00.0-
HSBBB0.3730.1620.426
HSHMC0.2570.1210.43
HSNLM + 2S0.2440.2010.252
HSBBB + 2S0.2840.0940.387
HSHMC + 2S0.2790.1110.427

Table 3. Evaluation Metrics of All Bundle Networks and Equations in the Training Region. Here we Use 2S for Two-step and EB for Error Bounds.

EquationMethodMedian REMedian ResidualMiscal. Area
ΛCDMFCNN0.3740.027-
ΛCDMBBB0.9860.1680.495
ΛCDMHMC0.31611.9790.444
ΛCDMNLM + 2S0.44910.5690.43
ΛCDMBBB + 2S0.36312.6720.495
ΛCDMHMC + 2S0.3112.0540.424
ΛCDMNLM + 2S + EB0.28714.7780.113
ΛCDMBBB + 2S + EB0.37510.6950.155
ΛCDMHMC + 2S + EB0.32911.4630.168
CPLFCNN0.0330.0-
CPLBBB0.3730.0130.357
CPLHMC0.0820.0120.152
CPLBBB + 2S0.590.0010.495
CPLHMC + 2S0.2750.0010.072
CPLBBB + 2S + EB0.9580.0010.216
CPLHMC + 2S + EB0.6480.0010.194
Quint.FCNN0.0220.016-
Quint.BBB0.1250.0890.349
Quint.HMC0.0060.0110.021
Quint.NLM + 2S0.1380.3350.334
Quint.BBB + 2S0.2210.130.377
Quint.HMC + 2S0.0380.030.096
HSFCNN0.3740.027-
HSBBB0.3740.1610.427
HSHMC0.2570.1210.43
HSNLM + 2S6.9990.5920.412
HSBBB + 2S2.1241.6160.452
HSHMC + 2S0.8850.9840.487

Table 4. Evaluation Metrics of All Bundle Networks and Equations in the OOD Region. Here we Use 2S for Two-step and EB for Error Bounds.

In the OOD region, results are more mixed. For Λ\LambdaCDM, the combination of two steps and error bounds significantly reduces miscalibration. However, in the CPL model, applying two steps increases miscalibration for BBB but reduces it for HMC; adding error bounds reverses this behavior—improving BBB but slightly worsening HMC. As shown in Figure 26, this is likely because FCNN closely tracks the ground truth in the OOD region, leading to narrow error bounds, while BBB and HMC tend to drift from the true function, which error bounds fail to capture accurately. For the more complex equations, Quintessence and HS, the benefits of the two-step method are less consistent, and no clear advantage is observed over the baseline approaches.

CPL Bundle Solutions (Error Bounds) plots

Figure 26. Examples Of CPL Bayesian Solutions Obtained Using The Bundle Solver With Error Bounds. Analytic Solutions Are Presented In Dotted Lines.

In Table 11 we can see how the Bayesian methods provide tighter intervals than deterministic nets. This is due to observations being more likely under the distributions generated by Bayesian methods. These uncertainties can be propagated through the approximated solution to obtain solution distributions that account for the uncertainties in the equation parameters.

Table 12 shows the level of agreement of our results with the literature. Most of our results are in an agreement of 1σ1\sigma or 2σ2\sigma. ΛCDM is an exception, where we observe higher disagreement due to the tight bounds obtained and discrepancy in the means.

Overall, our results demonstrate that leveraging error bounds within a two-step framework significantly enhances uncertainty quantification, particularly in calibration. However, the observed tradeoff between calibration and numerical stability suggests the need for further refinement, especially for highly complex equations such as HS.

Among Bayesian approaches, HMC generally achieves superior accuracy and calibration, but its computational cost is substantial. BBB, on the other hand, performs well when combined with our two-step procedure, offering a more computationally efficient and flexible alternative. There is considerable potential for improving BBB by adopting more expressive variational posterior distributions. Finally, NLM also demonstrates strong performance while being the least expensive method, though its applicability is limited by the underlying distributional assumptions.

References

Yashar Akrami, Renata Kallosh, Andrei Linde, and Valeri Vardanyan. The landscape, the swampland and the era of precision cosmology. Fortschritte der Physik, 67 (1-2):1800075, January 2019. ISSN 0015-8208, 1521-3978. doi: 10.1002/prop.201800075. URL http://arxiv.org/abs/1808.09440. arXiv:1808.09440 [astro-ph, physics:gr-qc, physics:hep-ph, physics:hep-th, physics:math-ph].

C. Armendariz-Picon, V. Mukhanov, and Paul J. Steinhardt. Dynamical Solution to the Problem of a Small Cosmological Constant and Late-Time Cosmic Acceleration. Physical Review Letters, 85(21):4438–4441, November 2000. doi: 10.1103/PhysRevLett.85.4438. URL https://link.aps.org/doi/10.1103/PhysRevLett.85.4438. Publisher: American Physical Society.

Eli Bingham, Jonathan P. Chen, Martin Jankowiak, Fritz Obermeyer, Neeraj Pradhan, Theofanis Karaletsos, Rohit Singh, Paul Szerlip, Paul H. Skrfsl, and Noah D. Goodman. Pyro: Deep Universal Probabilistic Programming, Oc-

Supplementary Material

Pablo Flores; Olga Graf; Pavlos Protopapas; Karim Pichara

Enforcing IC/BC

Two-dimensional Dirichlet Boundary Value Problem

In the Dirichlet problem, also known as first boundary value problem, the values of the solution function at the boundaries Γ\Gamma are known [6]. In this case the re-parameterization is

u~θ(x,y):=A(x,y)+x~(1−x~)y~(1−y~)uθ(x,y)\tilde{u}_{\theta}(x,y) := A(x,y) + \tilde{x}(1-\tilde{x})\tilde{y}(1-\tilde{y})u_{\theta}(x,y)

where

A(x,y)=(1−x~)u(x0,y)+x~u(x1,y)+(1−y~)(u(x,y0)−(1−x~)u(x0,y0)+x~u(x1,y0))+y~(u(x,y1)−(1−x~)u(x0,y1)+x~u(x1,y1))\begin{aligned} A(x,y) ={}& (1-\tilde{x})u(x_{0},y) + \tilde{x}u(x_{1},y) \\ &+ (1-\tilde{y})\left(u(x,y_{0})-(1-\tilde{x})u(x_{0},y_{0})+\tilde{x}u(x_{1},y_{0})\right) \\ &+ \tilde{y}\left(u(x,y_{1})-(1-\tilde{x})u(x_{0},y_{1})+\tilde{x}u(x_{1},y_{1})\right) \end{aligned}
x~=x−x0x1−x0\tilde{x} = \frac{x-x_{0}}{x_{1}-x_{0}}
y~=y−y0y1−y0\tilde{y} = \frac{y-y_{0}}{y_{1}-y_{0}}

(x0,y),(x1,y),(x,y0),(x,y1)∈Γ(x_{0},y),(x_{1},y),(x,y_{0}),(x,y_{1}) \in\Gamma

Error Bounds Computation

First Order Linear ODE with Constant Coefficient

A first-order linear ODE with constant coefficient has the general form

u′(t)+(λ+iω)u(t)=f(t)(29)u'(t) + (\lambda+i\omega)u(t)=f(t) \tag*{(29)}

As shown by Liu et al. [2022], the error of a PINN solution uθu_{\theta} to Eq. (29) can be bounded as

∣uθ(t)−u(t)∣≤εe−λt∫t0teλτ dτ(30)|u_{\theta}(t)-u(t)| \le\varepsilon e^{-\lambda t}\int_{t_{0}}^{t}e^{\lambda\tau}\,\mathrm{d}\tau \tag*{(30)}

when the initial conditions are satisfied i.e uθ(t0)=u(t0)u_{\theta}(t_{0})=u(t_{0}). Here, ε\varepsilon is an upper bound on the residuals

∣u′(t)+(λ+iω)u(t)−f(t)∣≤ε∀t∈I(31)|u'(t)+(\lambda+i\omega)u(t)-f(t)| \le\varepsilon\quad\forall t \in I \tag*{(31)}

where II can be any of the forms (t0,t)(t_{0},t), (t0,t](t_{0},t], (t0,∞)(t_{0},\infty).

Algorithm 1 Tight Error Bounds

Input: Domain I=[Tmin⁡,Tmax⁡]I = [T_{\min}, T_{\max}], residual network r(t)r(t), an expression for eP(t)e^{P(t)} we will call EP⁡(t)\operatorname{EP}(t), an expression for ∫Tmin⁡teP(τ)dτ\int_{T_{\min}}^{t} e^{P(\tau)}\mathrm{d}\tau we will call IntEP⁡(t)\operatorname{IntEP}(t), number of partitions NN, number of points in each partition KK.

Output: A set of times and error bounds at those times {ti,bi}i=1N\{t_i,b_i\}_{i=1}^{N}.

{ti}i=0N←Linspace⁡(Tmin⁡,Tmax⁡,N+1)\{t_i\}_{i=0}^{N} \gets\operatorname{Linspace}(T_{\min},T_{\max},N+1)

Initialize b0:=0b_0 := 0

for i←1…Ni \gets1 \ldots N do

    Ii←Linspace⁡(ti−1,ti,K)I_i \gets\operatorname{Linspace}(t_{i-1},t_i,K)

    εi←max⁡τ∈Ii∣r(τ)∣\varepsilon_i \gets\max_{\tau\in I_i}|r(\tau)|

    bi←bi−1+εiIntEP⁡(ti)−IntEP⁡(ti−1)EP⁡(ti)b_i \gets b_{i-1}+\varepsilon_i \frac{\operatorname{IntEP}(t_i)-\operatorname{IntEP}(t_{i-1})}{\operatorname{EP}(t_i)}     ▷ Implementation of Eq. (37)

end for

Note: Linspace⁡(a,b,n)\operatorname{Linspace}(a,b,n) returns an array of nn points equally spaced in the range [a,b][a,b].

First Order Linear ODE with Nonconstant Coefficient

A first-order linear ODE with nonconstant coefficient has the general form

u′(t)+(p(t)+iq(t))u(t)=f(t)(32)u'(t) + (p(t) + \mathrm{i}q(t))u(t) = f(t) \tag*{(32)}

As shown by Liu et al. [2022], when the initial conditions are satisfied, the error of a PINN solution uθu_\theta to Eq. (32) can be bounded as

∣uθ(t)−u(t)∣≤εe−P(t)∫t0teP(τ)dτ(33)|u_\theta(t)-u(t)| \le\varepsilon e^{-P(t)}\int_{t_0}^{t}e^{P(\tau)}\mathrm{d}\tau \tag*{(33)}

Where P(t)=∫t0tp(τ)dτP(t)=\int_{t_0}^{t}p(\tau)\mathrm{d}\tau, and ε\varepsilon is an upper bound on the residuals

∣u′(t)+(λ+iω)u(t)−f(t)∣≤ε∀t∈I(34)|u'(t)+(\lambda+\mathrm{i}\omega)u(t)-f(t)|\le\varepsilon\quad\forall t\in I \tag*{(34)}

where II can be any of the forms (t0,t)(t_0,t), (t0,t](t_0,t], (t0,∞)(t_0,\infty).

Tight Bounds Computation

In Section 2.6, we described how to obtain a bound using the global maximum residual ε\varepsilon in II. However, we can compute a tighter error bound by partitioning the domain I=I1⊎I2⊎⋯⊎InI=I_1\mathbin{\uplus}I_2\mathbin{\uplus}\cdots\mathbin{\uplus}I_n.

The tight bound for the first-order ODE with nonconstant coefficients turns out to be

∣uθ(t)−u(t)∣≤e−P(t)∫t0t∣rθ(τ)∣eP(τ)dτ(35)|u_\theta(t)-u(t)|\le e^{-P(t)}\int_{t_0}^{t}|r_\theta(\tau)|e^{P(\tau)}\mathrm{d}\tau \tag*{(35)}

To use the partitions IkI_k, we define the maximum local residual

εk:=max⁡τ∈Ik∣rθ(τ)∣(36)\varepsilon_k:=\max_{\tau\in I_k}|r_\theta(\tau)| \tag*{(36)}

and we compute the bounds as

∣uθ(t)−u(t)∣≤∑i=1nεie−P(t)∫τ=ti−1τ=tieP(τ)dτ(37)|u_\theta(t)-u(t)|\le\sum_{i=1}^{n}\varepsilon_i e^{-P(t)}\int_{\tau=t_{i-1}}^{\tau=t_i}e^{P(\tau)}\mathrm{d}\tau \tag*{(37)}

where tk=max⁡Ikt_k=\max I_k and tn=tt_n=t. Algorithm 1 in the Supplementary Material shows the implementation of tight bounds computation.

NLM Details

We now provide details on the NLM derivation presented in Section 4.1.1.

Σpost=(ΦθTΣΦθ+σprior−2I)−1(38)\Sigma_{\mathrm{post}}=\left(\Phi_\theta^{T}\Sigma\Phi_\theta+\sigma_{\mathrm{prior}}^{-2}I\right)^{-1} \tag*{(38)}
μpost=Σpost(Σ−1Φθ)Tuθdet(39)\mu_{\mathrm{post}}=\Sigma_{\mathrm{post}}\left(\Sigma^{-1}\Phi_\theta\right)^{T}u_{\theta_{\mathrm{det}}} \tag*{(39)}

we have written ΦθT\Phi_{\theta}^{T} and Σ\Sigma instead of ΦθT(xD,λD)\Phi_{\theta}^{T}(x_D,\lambda_D) and Σ(xD,λD)\Sigma(x_D,\lambda_D) respectively, to simplify notation. In Eq. (38) and Eq. (39),

xD=[x1,…,xN]Tx_D = [x_1,\ldots,x_N]^T
λD=[λ1,…,λN]T\lambda_D = [\lambda_1,\ldots,\lambda_N]^T
Σ(xD,λD)=diag⁡([σLike2(x1,λ1),…,σLike2(xM,λM)])\Sigma(x_D,\lambda_D)=\operatorname{diag}([\sigma_{\mathrm{Like}}^2(x_1,\lambda_1),\ldots,\sigma_{\mathrm{Like}}^2(x_M,\lambda_M)])

Computing Predictive Uncertainty

The computation of predictive uncertainty is straightforward as it is the standard deviation of the posterior predictive distribution. We can obtain it by applying the law of total variance

Var⁡(u∣x,λ,D)=Eθ∣D[Var⁡(u∣x,λ,θ)]+Var⁡θ∣D[E(u∣x,λ,θ)]=Eθ∣D[σLike2(x,λ)]+Var⁡θ∣D[uθ(x,λ)]≈σLike2(x,λ)+1M∑i=1M(uθi(x,λ)−u‾(x,λ))2(40)\begin{aligned} \operatorname{Var}(u|x,\lambda,D) &= \mathbb{E}_{\theta|D}[\operatorname{Var}(u|x,\lambda,\theta)] + \operatorname{Var}_{\theta|D}[\mathbb{E}(u|x,\lambda,\theta)] \\ &= \mathbb{E}_{\theta|D}[\sigma_{\mathrm{Like}}^2(x,\lambda)] + \operatorname{Var}_{\theta|D}[u_\theta(x,\lambda)] \\ &\approx\sigma_{\mathrm{Like}}^2(x,\lambda) + \frac{1}{M}\sum_{i=1}^{M}(u_{\theta_i}(x,\lambda)-\overline{u}(x,\lambda))^2 \tag*{(40)} \end{aligned}

where θi∼p(θ∣D)\theta_i \sim p(\theta|D) and u‾(x,λ)=1M∑i=1Muθi(x,λ)\overline{u}(x,\lambda)=\frac{1}{M}\sum_{i=1}^{M}u_{\theta_i}(x,\lambda) is the sample mean of the network outputs.

The latter approximation applies to BBB and NUTS; for BBB, the samples are taken from the variational posterior, and for NUTS, we use the samples of the true posterior the method yields. We have the analytical expression for NLM as shown in Eq. (38).

Statistical Analysis

It may also be the case that the observed values are a function ff of the solution. In that scenario, the dataset would become Of={(xi,μif,σif)∣xi∈Ω,μif∈R,σif∈R+}i=1Of\mathcal{O}^{f}=\{(x_i,\mu_i^{f},\sigma_i^{f})\mid x_i\in\Omega,\mu_i^{f}\in\mathbb{R},\sigma_i^{f}\in\mathbb{R}^{+}\}_{i=1}^{\mathcal{O}^{f}}, where μif\mu_i^{f} and σif\sigma_i^{f} are the mean and standard deviation of the observed function, respectively. Here ff is a function of the true solution and possibly other parameters λf\lambda^{f}, which can also be included in the set of parameters to be estimated. In a similar way to Eq. (24), we can compute the likelihood:

p(xi,μif,σif∣λ,λf)=∫Up(xi,μif,σif∣f(uλ(xi),λf))⋅p(uλ(xi)∣xi,λ) du(41)p(x_i,\mu_i^{f},\sigma_i^{f}|\lambda,\lambda^{f})=\int_{\mathcal{U}}p(x_i,\mu_i^{f},\sigma_i^{f}|f(u_\lambda(x_i),\lambda^{f}))\cdot p(u_\lambda(x_i)|x_i,\lambda)\,du \tag*{(41)}

Proofs

Proposition F.1. Given a differential equation of the form dudt=f(t)\frac{du}{dt}=f(t) for some function ff and an initial condition u(t0)=u0u(t_0)=u_0. The error of an approximation u^\hat{u} of uu is:

u^(τ)−u(τ)=∫t0τR(t)dt(42)\hat{u}(\tau)-u(\tau)=\int_{t_0}^{\tau}R(t)dt \tag*{(42)}

if the approximation satisfies the ICs, where R(t):=du^dt−f(t)R(t):=\frac{d\hat{u}}{dt}-f(t).

Proof. Let us take some approximate solution to the DE

u^(t)=u(t)+eu(t)\hat{u}(t)=u(t)+e_u(t)

where u^(t)\hat{u}(t) is the approximation, u(t)u(t) is the true solution and eu(t)e_u(t) is an error term. We can do the same for the first derivative of the solution. Using Newton’s notation:

u^˙(t)=u˙(t)+e˙u(t)(43)\dot{\hat{u}}(t)=\dot{u}(t)+\dot{e}_u(t) \tag*{(43)}

Using the definition of R(t)R(t)

R(t)=u^˙(t)+eu˙(t)−f(t)R(t)=\dot{\hat{u}}(t)+e_{\dot{u}}(t)-f(t)

Since u˙(t)\dot{u}(t) is the derivative of the true solution, u˙(t)−f(t)=0\dot{u}(t)-f(t)=0, thus R(t)=eu˙(t)R(t)=e_{\dot{u}}(t). Substituting that in Eq. (43), we get

R(t)=u^˙(t)−u˙(t)R(t)=\dot{\hat{u}}(t)-\dot{u}(t)

Integrating both sides with respect to tt:

∫t0τR(t) dt=∫t0τu^˙(t) dt−∫t0τu˙(t) dt\int_{t_0}^{\tau}R(t)\,dt=\int_{t_0}^{\tau}\dot{\hat{u}}(t)\,dt-\int_{t_0}^{\tau}\dot{u}(t)\,dt
∫t0τR(t) dt=u^(τ)−u^(t0)+u(t0)−u(τ)\int_{t_0}^{\tau}R(t)\,dt=\hat{u}(\tau)-\hat{u}(t_0)+u(t_0)-u(\tau)
∫t0τR(t) dt+eu(t0)=u^(τ)−u(τ)\int_{t_0}^{\tau}R(t)\,dt+e_u(t_0)=\hat{u}(\tau)-u(\tau)

If u(t0)=u^(t0)u(t_0)=\hat{u}(t_0), which is the case for our PINNs because we enforce ICs, then

u^(τ)−u(τ)=∫t0τR(t) dt\hat{u}(\tau)-u(\tau)=\int_{t_0}^{\tau}R(t)\,dt

Cosmology Models

CDM Error Bounds Computation

Error Bounds Computation Since this model is a first-order linear system with non-constant coefficients, it is possible to compute its error bounds. To compute the error bounds with Algorithm 1 we need an expression for eP(z)e^{P(z)} and ∫0zeP(s) ds\int_0^z e^{P(s)}\,ds. In the Λ\LambdaCDM model, we have p(z)=3/(1+z)p(z)=3/(1+z), thus

P(z)=∫0z−31+s ds=−3ln⁡(1+s)(44)P(z)=\int_0^z\frac{-3}{1+s}\,ds=-3\ln(1+s) \tag*{(44)}

then

eP(s)=e−3ln⁡(1+s)(45)e^{P(s)}=e^{-3\ln(1+s)} \tag*{(45)}
∫0zeP(s) ds=∫0ze−3ln⁡(1+s) ds=12−12(1+z)2(46)\int_0^z e^{P(s)}\,ds=\int_0^z e^{-3\ln(1+s)}\,ds=\frac{1}{2}-\frac{1}{2(1+z)^2} \tag*{(46)}

which are the inputs needed for Algorithm 1.

Parametric Dark Energy Error Bounds Computation

Once again, we need to find an expression for the inputs to Algorithm 1. CPL’s equation is also a first-order linear with a non-constant coefficient equation, so the procedure is the same as for Λ\LambdaCDM. First, we have

p(z)=−31+z(1+w0+w1z1+z)(47)p(z)=\frac{-3}{1+z}\left(1+w_0+\frac{w_1z}{1+z}\right) \tag*{(47)}
P(z)=∫0z−31+s(1+w0+w1s1+s) ds(48)P(z)=\int_0^z\frac{-3}{1+s}\left(1+w_0+\frac{w_1s}{1+s}\right)\,ds \tag*{(48)}
=−3((w0+w1+1)ln⁡(z+1)+w1z+1−w1)(49)=-3\left((w_0+w_1+1)\ln(z+1)+\frac{w_1}{z+1}-w_1\right) \tag*{(49)}

we can then simplify the expression eP(z)e^{P(z)} as

eP(z)=e−3((w0+w1+1)ln⁡(z+1)+w1z+1−w1)(50)e^{P(z)}=e^{-3((w_0+w_1+1)\ln(z+1)+\frac{w_1}{z+1}-w_1)} \tag*{(50)}
=−3((w0+w1+1)ln⁡(z+1)+w1z+1−w1)(51)=-3((w_0+w_1+1)\ln(z+1)+\frac{w_1}{z+1}-w_1) \tag*{(51)}
=(z+1)−3(w0+w1+1)e3w1zz+1(52)=(z+1)^{-3(w_0+w_1+1)}e^{\frac{3w_1z}{z+1}} \tag*{(52)}

Finally, the integral turns out to be

∫0zeP(s) ds=e3w1(s+1)1−3(w0+w1+1)E2−3(w0+w1+1)(3w1s+1)∣s=0s=z(53)\int_0^z e^{P(s)}\,\mathrm{d}s=e^{3w_1}(s+1)^{1-3(w_0+w_1+1)}E_{2-3(w_0+w_1+1)}\left(\frac{3w_1}{s+1}\right)\Bigg|_{s=0}^{s=z} \tag*{(53)}

Where En(x)E_n(x) is the exponential integral defined as En(x)=∫1∞e−xttn dtE_n(x)=\int_1^\infty\frac{e^{-xt}}{t^n}\,\mathrm{d}t. We used computational tools to find the antiderivative of eP(s)e^{P(s)}.

Quintessence

The Quintessence [10, 2, 17] (Zlatev et al. 1999) system of equations is

dxdN=−3x+62λy2+32x(1+x2−y2)\frac{dx}{dN}=-3x+\frac{\sqrt{6}}{2}\lambda y^2+\frac{3}{2}x(1+x^2-y^2)
dydN=62xyλ+32y(1+x2−y2)\frac{dy}{dN}=\frac{\sqrt{6}}{2}xy\lambda+\frac{3}{2}y(1+x^2-y^2)

with the following initial conditions

x0=0x_0=0
y0=1−Ωm,0ΛΩm,0Λ(1+z0)3+1−Ωm,0Λy_0=\sqrt{\frac{1-\Omega_{m,0}^{\Lambda}}{\Omega_{m,0}^{\Lambda}(1+z_0)^3+1-\Omega_{m,0}^{\Lambda}}}

where λ\lambda, Ωm,0Λ\Omega_{m,0}^{\Lambda} are parameters. Given a solution to the system, the Hubble parameter is computed as

H(z)=H0ΛΩm,0Λ(1+z)31−x2−y2H(z)=H_0^{\Lambda}\sqrt{\frac{\Omega_{m,0}^{\Lambda}(1+z)^3}{1-x^2-y^2}}

Here H0ΛH_0^{\Lambda} is also a parameter.

f(R)f(R) Gravity

The f(R)f(R) gravity [15, 7] model, which we refer to as HS for the authors Hu and Sawicki [2007] consists of a system of 5 equations.

dxdz=1z+1(−Ω−2v+x+4y+xv+x2)\frac{dx}{dz}=\frac{1}{z+1}(-\Omega-2v+x+4y+xv+x^2)
dydz=−1z+1(vxΓ−xy+4y−2yv)\frac{dy}{dz}=\frac{-1}{z+1}(vx\Gamma-xy+4y-2yv)
dvdz=−vz+1(xΓ+4−2v)\frac{dv}{dz}=\frac{-v}{z+1}(x\Gamma+4-2v)
dΩdz=Ωz+1(−1+2v+x)\frac{d\Omega}{dz}=\frac{\Omega}{z+1}(-1+2v+x)
drdz=−rΓxz+1\frac{dr}{dz}=-\frac{r\Gamma x}{z+1}

where

Γ(r)=(r+b)[(r+b)2−2b]4r\Gamma(r) = \frac{(r+b)\left[(r+b)^2-2b\right]}{4r}

The initial conditions are

x0=0x_0=0
y0=Ωm,0Λ(1+z0)3+2(1−Ωm,0Λ)2[Ωm,0Λ(1+z0)3+(1−Ωm,0Λ)]y_0=\frac{\Omega^\Lambda_{m,0}(1+z_0)^3+2(1-\Omega^\Lambda_{m,0})}{2\left[\Omega^\Lambda_{m,0}(1+z_0)^3+(1-\Omega^\Lambda_{m,0})\right]}
v0=Ωm,0Λ(1+z0)3+4(1−Ωm,0Λ)2[Ωm,0Λ(1+z0)3+(1−Ωm,0Λ)]v_0=\frac{\Omega^\Lambda_{m,0}(1+z_0)^3+4(1-\Omega^\Lambda_{m,0})}{2\left[\Omega^\Lambda_{m,0}(1+z_0)^3+(1-\Omega^\Lambda_{m,0})\right]}
Ω0=Ωm,0Λ(1+z0)3Ωm,0Λ(1+z0)3+(1−Ωm,0Λ)\Omega_0=\frac{\Omega^\Lambda_{m,0}(1+z_0)^3}{\Omega^\Lambda_{m,0}(1+z_0)^3+(1-\Omega^\Lambda_{m,0})}
r0=Ωm,0Λ(1+z0)3+4(1−Ωm,0Λ)1−Ωm,0Λr_0=\frac{\Omega^\Lambda_{m,0}(1+z_0)^3+4(1-\Omega^\Lambda_{m,0})}{1-\Omega^\Lambda_{m,0}}

where bb, Ωm,0Λ\Omega^\Lambda_{m,0} are the system parameters. Given a solution to the system, the Hubble parameter is computed as

H(z)=H0Λr2v(1−Ωm,0Λ)H(z)=H^\Lambda_0\sqrt{\frac{r}{2v}(1-\Omega^\Lambda_{m,0})}

Here H0ΛH^\Lambda_0 is also a parameter.

Implementation Details

As described in § 6, we followed the implementation from [11] to train the FCNNs in the first step. Architecture and hyperparameter details are shown in Table 5 to Table 8. NLM shares FCNN details since it builds on FCNN.

EquationInput Dim.Output Dim.Hidden UnitsActivationIterationsSamples per Dim. in a BatchLearning Rate
Λ\LambdaCDM11(32, 32)Tanh100,000640.001
CPL12(32, 32)Tanh100,0001280.001
Quintessence11(32, 32)Tanh100,000320.001
HS11(32, 32)Tanh600,000320.001

Table 5. Implementation Details For FCNN.

EquationInput Dim.Output Dim.Hidden UnitsActivationPosterior SamplesTune SamplesSamples per Dim.Prior Std.Likelihood Std.
Λ\LambdaCDM11(32, 32)Tanh10,0001,0003210.1
CPL12(32, 32)Tanh10,0001,00012810.01
Quintessence11(32, 32)Tanh10,0001,0003210.005
HS11(32, 32)Tanh10,0001,0003210.005

Table 8. Implementation Details For HMC.

We used Pytorch [Paszke et al., 2019] for FCNNs, Pyro [3] for BNNs, and Neurodiffeq [12] for aiding in PINNs training.

EquationSamples per Dim.Likelihood Std.
Λ\LambdaCDM1000.1
Quintessence320.005
HS320.005

Table 6. Implementation Details For NLM.

EquationInput Dim.Output Dim.Hidden UnitsActivationIterationsSamples per Dim.Prior Std.Learning RateLikelihood Std.
Λ\LambdaCDM11(32, 32)Tanh20,0006410.0010.1
CPL12(32, 32)Tanh20,00012810.0010.01
Quintessence11(32, 32)Tanh20,0003210.0010.005
HS11(32, 32)Tanh20,0003210.0010.005

Table 7. Implementation Details For BBB.

Exploration of Solutions, Residuals and Errors

To analyze the behavior of residuals and errors in PINNs, we conducted a systematic data collection process. We trained N deterministic PINNs for I iterations to solve each of the cosmological models under study. Throughout training, we recorded solutions and residuals every 10 iterations, evaluating them on a fixed, equidistant set of T time steps. Detailed specifications for each cosmological model are provided in Table 10.

EquationNumber of NN (N)Iterations (I)Time Steps (T)
Λ\LambdaCDM1000100050
CPL1000100050
Quintessence1500100050
HS200500050

Table 10. Implementation Details of Appendix I.

Distribution of Solutions After collecting the data, we generated histograms of the solutions throughout training, as shown in Figure 4 to Figure 12. For reference, we overlaid Gaussian distributions with the sample mean and variance in orange. The distribution of solutions varies across cosmological models, with Λ\LambdaCDM and CPL being the closest to Gaussian, while HS exhibits noticeable bimodality.

Residuals and Errors Relationship Understanding the relationship between solution errors and residuals is crucial for assessing whether residuals can serve as a reliable source of calibrated uncertainties. Ideally, a perfect correlation between residuals and errors would indicate that residuals effectively capture uncertainty.

In Figure 13 to Figure 16, we plot the absolute, normalized values of relative errors against their corresponding residuals for all cosmological models. The results indicate no clear correlation; in fact, the trend often shows higher errors associated with lower residuals. This behavior aligns with the implicit bias described by Wang et al. [2022], where PINNs exhibit lower convergence rates near initial conditions.

Distribution of $x_{\mathrm{DE}}(z)$ from CPL across iterations and time values

Figure 5. Distribution of xDE(z)x_{\mathrm{DE}}(z) from CPL. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Distribution of $x(N)$ from Quintessence

Figure 6. Distribution of x(N)x(N) from Quintessence. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Grid of histograms showing distributions of $y(N)$ for Quintessence across iterations 0, 250, 500, 750, and 1000 and times $t=0.0$, $0.25$, $0.5$, $0.75$, and $1.0$

Figure 7. Distribution of y(N)y(N) from Quintessence. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Distribution histograms of $x(z)$ across iterations and values of $t$

Figure 8. Distribution of x(z)x(z) from HS. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Distribution of $y(z)$ from HS

Figure 9. Distribution of y(z)y(z) from HS. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Distribution of $v(z)$ from HS, with histograms and Gaussian fits across iterations and values of $t$

Figure 10. Distribution of v(z)v(z) from HS. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Distribution of $\Omega(z)$ from HS

Figure 11. Distribution of Ω(z)\Omega(z) from HS. Samples were collected as described in I. Orange lines show a Gaussian distribution with sample mean and variance.

Relationship between solution errors and residuals from CPL

Figure 14. Relationship Between Solution Errors and Residuals from CPL. Samples were collected as described in I.

Relationship between solution errors and residuals from Quintessence

Figure 15. Relationship Between Solution Errors and Residuals from Quintessence. Samples were collected as described in I. Colors differentiate variables in the system.

Aditional Results

ΛCDM Bayesian solutions

Figure 17. Λ\LambdaCDM Bayesian Solutions. The Analytic Solution Is Presented In Dotted Lines.

ΛCDM Bayesian solutions with error bounds

Figure 18. Λ\LambdaCDM Bayesian Solutions With Error Bounds. The Analytic Solution Is Presented In Dotted Lines.

CPL Bayesian solutions

Figure 19. CPL Bayesian Solutions. The Analytic Solution Is Presented In Dotted Lines.

CPL Forward Solutions (Error Bounds)

Figure 20. CPL Bayesian Solutions With Error Bounds. The Analytic Solution Is Presented In Dotted Lines.

Quintessence Forward Solutions (Residual Likelihood)

Figure 21. Quintessence Bayesian Solutions Residual Likelihood. The Numerical Solution Is Presented In Dotted Lines.

Quintessence Forward Solutions

Figure 22. Quintessence Bayesian Solutions. The Numerical Solution Is Presented In Dotted Lines.

HS Forward Solutions (Residual Likelihood) for BBB and HMC

Figure 23. HS Bayesian solutions Residual Likelihood. The Numerical Solution Is Presented In Dotted Lines.

HS Forward Solutions plots for BBB, NLM, and HMC

Figure 24. HS Bayesian solutions. The Numerical Solution Is Presented In Dotted Lines.

CPL Bundle Solutions: four rows of plots for FCNN, BBB, and HMC, with axes labeled $x_{\mathrm{DE}}$ and $z$, and legend entries $w_1=-8.0$, $w_1=-6.11$, $w_1=-4.11$, $w_1=-2.22$, $w_1=-0.22$, and Training Region

Figure 25. Examples Of CPL Bayesian Solutions Obtained Using The Bundle Solver. Analytic Solutions Are Presented In Dotted Lines.

Quintessence Bundle Solutions for $\Omega_{m,0}=0.1$ plots

Figure 27. Examples of Quintessence Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.1\Omega_{m,0}=0.1. Numerical Solutions Are Presented In Dotted Lines.

Quintessence Bundle Solutions for $\Omega_{m,0}=0.2$

Figure 28. Examples of Quintessence Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.2\Omega_{m,0}=0.2. Numerical Solutions Are Presented In Dotted Lines.

Quintessence Bundle Solutions for $\Omega_{m,0}=0.3$

Figure 29. Examples of Quintessence Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.3\Omega_{m,0}=0.3. Numerical Solutions Are Presented In Dotted Lines.

Quintessence Bundle Solutions for $\Omega_{m,0}=0.4$

Figure 30. Examples of Quintessence Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.4\Omega_{m,0}=0.4. Numerical Solutions Are Presented In Dotted Lines.

Examples of HS Bayesian solutions obtained using the Bundle Solver for the parameter value $\Omega_{m,0} = 0.1$

Figure 31. Examples Of HS Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.1\Omega_{m,0} = 0.1. Numerical Solutions Are Presented In Dotted Lines.

HS Bundle Solutions for $\Omega_{m,0}=0.2$

Figure 32. Examples Of HS Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.2\Omega_{m,0} = 0.2. Numerical Solutions Are Presented In Dotted Lines.

HS Bundle Solutions for $\Omega_{m,0}=0.3$

Figure 33. Examples Of HS Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.3\Omega_{m,0}=0.3. Numerical Solutions Are Presented In Dotted Lines.

HS Bundle Solutions for $\Omega_{m,0}=0.4$, showing FCNN, NLM, BBB, and HMC panels with legend values $\Omega_{m,0}=0.0$, $\Omega_{m,0}=0.24$, $\Omega_{m,0}=0.49$, $\Omega_{m,0}=0.75$, and $\Omega_{m,0}=1.0$, and a Training Region

Figure 34. Examples Of HS Bayesian Solutions Obtained Using The Bundle Solver For The Parameter Value Ωm,0=0.4\Omega_{m,0}=0.4. Numerical Solutions Are Presented In Dotted Lines.

EquationMethodMREMean ResidualMiscal. AreaRMS Cal.MA Cal.SharpnessNLLCRPSCheckIntervalAcc. MAEAcc. RMSEAcc. MDAEAcc. MARPD
ΛCDMFCNN0.2160.0------------
ΛCDMBBB0.9550.1540.4910.5660.4860.10127843.20.01910.01208.42620.07328.17212.83187.161
ΛCDMHMC0.3613.1270.4910.5660.4860.516319.71711.2145.612113.0711.45319.0392.68749.016
ΛCDMNLM + 2S0.2266.5440.2240.2580.2220.546247.8629.7434.876298.2939.93917.9170.02332.164
ΛCDMBBB + 2S0.293.8070.1850.2220.1830.31100.19111.2925.648116.15911.4119.6620.45941.833
ΛCDMHMC + 2S0.233.8420.2030.2350.2010.634135.0589.7394.87597.6619.97717.8780.07532.568
ΛCDMNLM + 2S + EB0.2216.6230.1480.1860.14717.23−0.4725.9483.00425.5029.79217.7370.00531.357
ΛCDMBBB + 2S + EB0.2663.690.1140.1370.11317.2652.2166.9583.51330.45811.06619.1780.85538.671
ΛCDMHMC + 2S + EB0.2223.8410.1110.1380.1117.251−0.2085.9533.00625.6279.78817.7070.05531.387
CPLFCNN0.0020.0------------
CPLBBB0.0080.0140.3590.3940.3550.023−3.1960.0060.0030.0360.0060.0110.0010.766
CPLHMC0.0090.0140.2680.2880.2660.016−4.4580.0040.0020.020.0070.0120.00.905
CPLBBB + 2S0.0360.0260.1030.1140.1020.019−2.2840.0210.010.1230.0260.0430.0043.444
CPLHMC + 2S0.0160.0190.2670.30.2650.034−3.870.0080.0040.0410.0120.020.0011.592
CPLBBB + 2S + EB0.0210.0150.1150.1370.1140.007−2.3760.0130.0060.0930.0150.0270.0011.993
CPLHMC + 2S + EB0.1580.2570.3890.4430.3850.075.180.0930.0470.70.110.2920.010.464
QuintessenceFCNN0.0440.054------------
QuintessenceBBB0.8380.1090.4820.5530.4770.025163.7560.1270.0641.1130.140.1710.135nan
QuintessenceHMC0.060.070.1760.2040.1740.01−4.0340.0130.0060.080.0150.030.0nan
QuintessenceNLM + 2S0.0730.2570.190.230.1880.015−3.3170.0170.0080.1090.0210.0410.09.054
QuintessenceBBB + 2S0.2010.1470.1280.1810.1260.00440.0680.0360.0180.3520.0380.0680.002nan
QuintessenceHMC + 2S0.0720.0840.1240.1420.1230.016−4.9480.0140.0070.0680.0190.0360.0nan
HSFCNN0.2160.0------------
HSBBB1.2240.2150.4870.560.4820.0132094.8830.3240.1623.2840.330.3780.347nan
HSHMC0.2680.1280.440.5120.4360.013723.2930.1380.0691.3260.1440.1690.148nan
HSNLM + 2S0.3530.3240.1520.1840.1510.409−1.2790.1490.0750.8620.1970.3410.01235.335
HSBBB + 2S0.1580.3260.2190.2530.2160.01259.5060.0280.0140.2390.0310.0680.007nan
HSHMC + 2S0.2620.2120.2650.3040.2630.02190.2840.0240.0120.1820.0280.0660.006nan

Table 14. Metrics of the NN Forward Solutions.

References

  1. [1]Yashar Akrami, Renata Kallosh, Andrei Linde, and Valeri Vardanyan. The landscape, the swampland and the era of precision cosmology. Fortschritte der Physik, 67 (1-2):1800075, January 2019. ISSN 0015-8208, 1521-3978. doi: 10.1002/prop.201800075. URL http://arxiv.org/abs/1808.09440. arXiv:1808.09440 [astro-ph, physics:gr-qc, physics:hep-ph, physics:hep-th, physics:math-ph].
  2. [2]C. Armendariz-Picon, V. Mukhanov, and Paul J. Steinhardt. Dynamical Solution to the Problem of a Small Cosmological Constant and Late-Time Cosmic Acceleration. Physical Review Letters, 85(21):4438–4441, November 2000. doi: 10.1103/PhysRevLett.85.4438. URL https://link.aps.org/doi/10.1103/PhysRevLett.85.4438. Publisher: American Physical Society.
  3. [3]Eli Bingham, Jonathan P. Chen, Martin Jankowiak, Fritz Obermeyer, Neeraj Pradhan, Theofianis Karaletsos, Rohit Singh, Paul Szerlip, Paul Horsfall, and Noah D. Goodman. Pyro: Deep Universal Probabilistic Programming, Oc- tober 2018. URL http://arxiv.org/abs/1810.09538. arXiv:1810.09538 [cs, stat].
  4. [4]Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, July 2006. ISBN 978-0-387-31073-2.
  5. [5]Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. 32nd International Conference on Machine Learning, ICML 2015, 2:1613–1622, 2015. arXiv: 1505.05424 ISBN: 9781510810587.arxiv.org/abs/1505.05424
  6. [6]William E Boyce and Richard C DiPrima. Elementary differential equations and boundary value problems. John Wiley & Sons, Chichester, England, 9 edition, October 2008.
  7. [7]H. A. Buchdahl. Non-Linear Lagrangians and Cosmological Theory. Monthly Notices of the Royal Astronomical Society, 150(1):1–8, September 1970. ISSN 0035-8711. doi: 10.1093/mnras/150.1.1. URL https://doi.org/10.1093/mnras/150.1.1.
  8. [8]Shengze Cai, Zhiping Mao, Zhicheng Wang, Minglang Yin, and George Em Karniadakis. Physics-informed neural networks (PINNs) for fluid mechanics: a review. Acta Mechanica Sinica, 37(12):1727–1738, December 2021a. ISSN 1614-3116. doi: 10.1007/s10409-021-01148-1. URL https://doi.org/10.1007/s10409-021-01148-1.
  9. [9]Shengze Cai, Zhicheng Wang, Sifan Wang, Paris Perdikaris, and George Em Karniadakis. Physics-Informed Neural Networks for Heat Transfer Problems. Journal of Heat Transfer, 143(6), April 2021b. ISSN 0022-1481. doi: 10.1115/1.4050542. URL https://doi.org/10.1115/1.4050542.
  10. [10]R. R. Caldwell, Rahul Dave, and Paul J. Steinhardt. Cosmological Imprint of an Energy Component with General Equation of State. Physical Review Letters, 80(8):1582–1585, February 1998. doi: 10.1103/PhysRevLett.80.1582. URL https://link.aps.org/doi/10.1103/PhysRevLett.80.1582. Publisher: American Physical Society.DOI
  11. [11]Augusto T. Chantada, Susana J. Landau, Pavlos Protopapas, Claudia G. Scóccola, and Cecilia Garraffo. Cosmological informed neural networks to solve the background dynamics of the Universe. May 2022. URL http://arxiv.org/abs/2205.02945. arXiv:2205.02945 [astro-ph, physics:gr-qc, physics:hep-ph].
  12. [12]Feiyu Chen, David Sondak, Pavlos Protopapas, Marios Mattheakis, Shuheng Liu, Devansh Agarwal, and Marco Di Giovanni. NeuroDiffEq: A Python package for solving differential equations with neural networks. Journal of Open Source Software, 5(46):1931, February 2020. ISSN 2475-9066. doi: 10.21105/joss.01931. URL https://joss.theoj.org/papers/10.21105/joss.01931.
  13. [13]Michel Chevallier and David Polarski. Accelerating universes with scaling dark matter. International Journal of Modern Physics D, 10(02):213–223, April 2001. ISSN 0218-2718. doi: 10.1142/S0218271801000822. URL https://www.worldscientific.com/doi/abs/10.1142/S0218271801000822. Publisher: World Scientific Publishing Co.DOI
  14. [14]Youngseog Chung, Ian Char, Han Guo, Jeff Schneider, and Willie Neiswanger. Uncertainty Toolbox: an Open-Source Library for Assessing, Visualizing, and Improving Uncertainty Quantification, September 2021. URL http://arxiv.org/abs/2109.10254. arXiv:2109.10254 [cs, stat].
  15. [15]Timothy Clifton, Pedro G. Ferreira, Antonio Padilla, and Constantinos Skordis. Modified gravity and cosmology. Physics Reports, 513(1):1–189, March 2012. ISSN 0370-1573. doi: 10.1016/j.physrep.2012.01.001. URL https://www.sciencedirect.com/science/article/pii/S0370157312000105.DOI
  16. [16]Zhang Cong, Zhang Han, Yuan Shuo, Liu Siqi, Zhang Tong-Jie, and Sun Yan-Chun. Four new observational H(z) data from luminous red galaxies in the Sloan Digital Sky Survey data release seven. Research in Astronomy and Astrophysics, 14(10):1221, October 2014. ISSN 1674-4527. doi: 10.1088/1674-4527/14/10/002. URL https://dx.doi.org/10.1088/1674-4527/14/10/002.
  17. [17]Edmund J. Copeland, Andrew R. Liddle, and David Wands. Exponential potentials and cosmological scaling solutions. Physical Review D, 57(8):4686–4690, April 1998. doi: 10.1103/PhysRevD.57.4686. URL https://link.aps.org/doi/10.1103/PhysRevD.57.4686. Publisher: American Physical Society.DOI
  18. [18]Rocco D’Agostino and Rafael C. Nunes. Measurements of $H_0$ in modified gravity theories: The role of lensed quasars in the late-time Universe. Physical Review D, 101(10):103505, May 2020. ISSN 2470-0010, 2470-0029. doi: 10.1103/PhysRevD.101.103505. URL http://arxiv.org/abs/2002.06381. arXiv:2002.06381 [astro-ph, physics:gr-qc, physics:hep-ph].
  19. [19]Arka Daw, Jie Bu, Sifan Wang, Paris Perdikaris, and Anuj Karpatne. Mitigating Propagation Failures in PINNs using Evolutionary Sampling, October 2022. URL http://arxiv.org/abs/2207.02338. arXiv:2207.02338 [cs].

Paper details

vlnc:W_y9e8gye3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents