Pith. sign in

REVIEW 3 major objections 5 minor 13 references

Physics-Informed Neural Networks for the Relativistic Burgers Equation in the Exterior of a Schwarzschild Black Hole

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims its PINN architecture, with a shock-aware block and a Godunov-inspired loss, reproduces both smooth and shock solutions of the relativistic Burgers equation in curved spacetime.

desk verdict New shock-aware PINN architecture for a curved-spacetime Burgers model, with honest accuracy caveats, but a wrong Godunov flux branch undercuts the general robustness claim. read the letter →

arxiv 2506.00951 v2 pith:LXN5EMUJ submitted 2025-06-01 math.NA cs.NAgr-qc

Shuyang Xiang This is my paper · ORCID
classification math.NAcs.NAgr-qc MSC 65M7068T0735L6583C57
keywords physics-informedneuralnetworksrelativisticBurgersequationSchwarzschildspacetimeshockwavesGodunovfluxconservationlawsmesh-freemethodsgeneralrelativity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces a physics-informed neural network (PINN) for the relativistic Burgers equation on the exterior of a Schwarzschild black hole. The central claim is that a network built from a smooth fully connected block plus a trainable 'shock-aware jump block,' trained with a Godunov-inspired residual, can reproduce both smooth steady-state flows and shock waves, including a shock that moves in time. If this claim holds, it would show that a mesh-free machine-learning solver can handle discontinuities in curved spacetime, offering an alternative to traditional finite-volume and finite-difference schemes for relativistic fluid problems.

What carries the argument

The central object is the shock-aware jump block combined with the Godunov-inspired loss residual. The jump block predicts the shock position $r_s(t)$ from time $t$ via a sigmoid-scaled network, then forms a differentiable indicator $h(t,r)=\sigma(k(r-r_s(t)))$ whose sharpness $k$ is learned, allowing the output to be nearly discontinuous without losing automatic differentiation. The residual term replaces the direct PDE residual with a first-order Godunov flux approximation $F_G(v_L,v_R)$ of $\partial_r f(v)$, with $f(v)=\frac{v^2-1}{2(1-2M/r)}$ and shock speed $s=\frac{v_L+v_R}{2}(1-2M/r)$, which encodes the jump condition in the loss. The full loss combines this with $L^2$ terms for initial and boundary data, and is trained in three stages (warm start, soft physics, full physics) using L-BFGS.

What would settle it

Run the trained model on a Riemann problem with initial data $v_L>0$, $v_R<0$, and a shock whose exact speed is negative; if the flux formula's branch for that case returns $\min\{f(v_L),f(v_R)\}$, the learned solution will converge to a different conservation law than the true entropy solution, so the predicted wave speed will not match a high-resolution finite-volume reference.

Watch

Extended reading notes

Core claim

The paper's contribution is a PINN architecture that is designed to approximate weak solutions, not just smooth ones, to equation (1.1): $\partial_t \left( \frac{v}{(1-2M/r)^2} \right) + \partial_r \left( \frac{v^2-1}{2(1-2M/r)} \right) = 0$ on $(2M,\infty)$. The network output is the sum $v_\theta(t,r)=N_{\text{smooth}}(t,r)+N_{\text{shock}}(t,r)$, where the shock block uses a sigmoid-smoothed Heaviside centered at a predicted shock location $r_s(t)$ with trainable sharpness $k$. The loss residual is computed with a Godunov flux $F_G(v_L,v_R)$, so the Rankine-Hugoniot condition is built into training. The authors validate the method on three test cases: a steady state, a stationary shock, and an initial discontinuity that produces a moving shock, all on the Schwarzschild exterior.

Load-bearing premise

The load-bearing premise is that the numerical flux formula in the loss function returns the correct shock speed for every pair of left and right states, including data with $v_L>0>v_R$ and a left-moving shock, a case that the paper's experiments never exercise.

Editorial extensions

If this is right

  • The method offers a mesh-free route to simulating relativistic conservation laws in curved spacetimes, which could be useful near the event horizon where mesh generation is awkward.
  • The same architecture, a smooth block plus a trainable jump block, can be carried over to other hyperbolic systems on curved backgrounds by replacing the flux function.
  • The three-stage training strategy gives a concrete recipe for training PINNs on problems with discontinuities without vanishing or exploding gradients.
  • The paper itself states that the accuracy is not competitive with traditional methods such as the Glimm scheme, so the near-term value is flexibility and ease of implementation rather than precision.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The Godunov flux formula in Section 3.2 has a branch for Riemann data with $v_L>0>v_R$ and negative shock speed that returns $\min\{f(v_L),f(v_R)\}$ instead of $f(v_R)$; the experiments never test that branch, so an immediate test would run that Riemann problem and compare the learned wave speed with a high-resolution reference.
  • The shock block predicts a single shock location; extending it to multiple shocks or to a spatially varying shock curve would let the method handle more complex wave patterns in relativistic fluid dynamics.
  • If the flux-branch issue is corrected, the same architecture could be applied to the relativistic Euler equations on Schwarzschild or Kerr backgrounds, where the flux is vector-valued and the jump conditions are coupled.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript introduces a physics-informed neural network (PINN) for the relativistic Burgers equation on the exterior of a Schwarzschild black hole, equation (1.1). The proposed architecture combines a standard fully connected network with a 'shock-aware jump block' that uses a trainable shock location and a smoothed Heaviside function. A custom residual is defined using a Godunov-inspired numerical flux, and a three-phase training strategy is described. Numerical experiments are presented for a smooth steady state, a steady shock, and a moving shock, and the authors claim the network reproduces both smooth and discontinuous solutions.

Significance. If fully validated, the approach would offer a mesh-free alternative for simulating relativistic fluid flows with shocks in curved spacetimes, which is of interest for black hole accretion and gravitational wave applications. The paper is honest about the fact that the accuracy is far below traditional schemes such as the Glimm method. The main strengths are the concrete shock-aware architecture and the staged training procedure. However, as it stands, the paper is a qualitative demonstration rather than a rigorous numerical study.

major comments (3)
  1. [Section 3.2] The Godunov flux formula F_G(v_L, v_R) is incorrect for the branch v_L > v_R with s < 0. For a convex flux f(v) = (v^2 - 1)/(2(1 - 2M/r)), the exact Godunov flux for a left-moving shock is f(v_R), but the formula returns min{f(v_L), f(v_R)}. For Riemann data with v_L > 0 > v_R and v_L + v_R < 0, this gives f(v_L), which is the wrong upwind value. This residual would drive the network toward an effective conservation law different from (1.1). The numerical experiments in Section 4 never exercise this branch: the steady-shock test has s = 0, and the moving-shock test has both states of the same sign, so the error is hidden. Consequently, the claim that the architecture 'handles well shock discontinuities' is not supported for general admissible data.
  2. [Section 4] The paper reports only visual agreement with the exact solutions in Figures 2–4 and provides no error metrics, no comparison against a reference solution (e.g., the Glimm scheme cited as [8]), and no convergence study. The claim of reproducing discontinuous solutions is therefore not quantitatively supported. The authors should report, at minimum, the L1 or L2 error in the smooth regions and the shock position as a function of time for the moving-shock test, together with a collocation-convergence study.
  3. [Sections 3.2–3.3] The boundary condition v_bnd(t) that appears in the loss (3.4) is never specified; the paper only states that boundary conditions are applied at both ends of the truncated domain. Additionally, the training schedule for the loss weights λ_eqn, λ_ini, λ_bnd is given only qualitatively ('small', then 1.0). Without these details, the experiments cannot be reproduced or independently evaluated.
minor comments (5)
  1. [Section 5] The phrase 'pros and cos' should be 'pros and cons'.
  2. [Section 2.1, Eq. (2.4)] The second branch of the piecewise definition says 'r0 > 0', which should likely be 'r > r0'.
  3. [Section 3.1] The notation 'σ net(t)' should be 'σ_net(t)' for readability.
  4. [Section 3.2] The word 'Rankie-Hugoniot' is a typo for 'Rankine–Hugoniot'.
  5. [Section 2.1, Theorem 2.1] Theorem 2.1 is presented without proof and is attributed to [9]; it should be clearly marked as a quoted theorem from that reference rather than stated as a new result.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PINN is trained against the PDE residual and validated on analytically derived solutions; self-citations appear only as background.

full rationale

The paper's central claim is that a specific PINN architecture with a shock-aware jump block and a Godunov-inspired residual can reproduce smooth and discontinuous solutions of the relativistic Burgers equation on a Schwarzschild exterior. The derivation chain is not circular: the loss function in Eq. (3.3)-(3.4) is built from the PDE residual (1.1) itself, with time derivatives obtained by automatic differentiation and spatial flux derivatives evaluated through an explicit numerical flux. The network output is not defined in terms of the target solution, and no fitted quantity is later relabeled as a prediction. The exact steady-state and shock solutions used for validation are derived directly from the equation in Section 2.1, e.g., formulas (2.2) and (2.4), and do not depend on the PINN parameters. References [8] and [9] include the author, but they are used as background on existence theory and previous numerical methods, not as evidence that the PINN works; the numerical experiments compare against explicit formulas and residual minimization rather than against the conclusions of those citations. No uniqueness theorem is imported from the authors to force the architecture choice, and no definitional equivalence between an input and an output conceals a circular step. The correctness issue raised in the reader's take concerning the Godunov flux formula for certain Riemann data is a mathematical accuracy concern, not a circularity concern, because the residual is still computed from an attempted discretization of the governing equation rather than from the desired answer.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on the cited existence theory for the PDE, on the exact steady solutions used as test cases, and on a custom numerical flux that is partially incorrect. The free parameters are training and discretization choices.

free parameters (4)
  • loss weights λ_eqn, λ_ini, λ_bnd = staged: 0, <<1, then 1.0
    Chosen by hand in the three-phase training strategy of Section 3.3.
  • slope limiter ε = 1e-5
    Small constant used to define left and right interface values in the Godunov residual (Section 3.2).
  • sharpness parameter k = trainable
    Controls the steepness of the smoothed Heaviside in the shock block; adjusted during training (Section 3.1).
  • truncation offset ε for r_min = unspecified
    The domain is truncated to r_min = 2M + ε to avoid the horizon singularity, but ε is not given a value (Section 3.2).
assumptions (5)
  • domain assumption Existence of weak solutions with locally finite total variation for BV initial data (Theorem 2.1)
    Invoked as background from [9] (LeFloch and Xiang) without proof in this paper.
  • standard math Steady-state and steady-shock solutions (2.2), (2.4) are exact solutions to (1.1)
    Used as initial conditions and implicit reference for validation; they are derived from the PDE.
  • ad hoc to paper Godunov flux formula in Section 3.2 is the correct numerical flux for (1.1)
    The residual is defined through this flux; the formula is incorrect for some Riemann data, e.g., v_L > 0 > v_R with negative shock speed.
  • domain assumption Domain truncation with boundary conditions at r_min and r_max yields a well-posed problem
    The unbounded domain is truncated to [r_min, r_max], but the boundary values v_bnd(t) are never specified.
  • ad hoc to paper Shock trajectory can be represented by a single sigmoid location r_s(t)
    The jump block models one discontinuity with a learned position; not valid for multiple or interacting shocks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Informed Neural Networks for the Relativistic Burgers Equation in the Exterior of a Schwarzschild Black Hole." pith.science (2026). https://pith.science/paper/LXN5EMUJ

@misc{pith2026250600951,
  author       = {Pith},
  title        = {Pith review of: Physics-Informed Neural Networks for the Relativistic Burgers Equation in the Exterior of a Schwarzschild Black Hole},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LXN5EMUJ}},
  note         = {Machine review of arXiv:2506.00951}
}
read the original abstract

We introduce a Physics-Informed Neural Networks(PINN) to solve a relativistic Burgers equation in the exterior domain of a Schwarzschild black hole. Our main contribution is a PINN architecture that is able to simulate shock wave formations in such curved spacetime, by training a shock-aware network block and introducing a Godunov-inspired residuals in the loss function. We validate our method with numerical experiments with different kinds of initial conditions. We show its ability to reproduce both smooth and discontinuous solutions in the context of general relativity.

Figures

Figures reproduced from arXiv: 2506.00951 by the authors.

Figure 1
Figure 1. Steady State solutions Theorem 2.1 (existence theory for the relativistic Burgers model) Consider the relativistic Burgers equation (1.1) posed on the outer domain of a Schwarzschild black hole with mass M. Given a function v0 : (2M, +∞) → (−1, 1) with locally finite total variation on (2M, +∞), that is, ∀a, b ∈ (2M,∞), TV[a,b](v) < ∞, then we have: there exists a corresponding weak solution to v = v(t, r) with loca… view at source ↗
Figure 2
Figure 2. Steady State solution We next validate the model in a setting of steady shock with the initial condition v0(r) =    q 1 − 3 4 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Steady Shock solution Finally, we evaluate the model an initial discontinuity which is not steady 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Shock solution the proposed PINN framework, enhanced with an adaptive shock modeling com￾ponent, is capable of learning both smooth and discontinuous solutions. 5 Discussion Our relativistic PINN to model compressible fluid flows in the Schwarzschild spacetime shows pr…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

  1. [8]

    LeFloch and Shuyang Xiang

    Philippe G. LeFloch and Shuyang Xiang. A numerical study of the rel- ativistic burgers and euler equations on a schwarzschild black hole exte- rior.Communications in Applied Mathematics and Computational Science, 13(2):271–301, 2018

    work page 2018

  2. [1]

    Amorim, P

    P. Amorim, P. G. LeFloch, and B. Okutmustur. Finite volume schemes on lorentzian manifolds.Communications in Mathematical Sciences, 6(4):1059–1086, 2008

    work page 2008

  3. [2]

    Matania Ben-Artzi, Joseph Falcovitz, and Philippe G. LeFloch. Hyper- bolic conservation laws on the sphere: A geometry-compatible finite volume scheme.Journal of Computational Physics, 228(16):5650–5668, 2009

    work page 2009

  4. [3]

    Solutions in the large for nonlinear hyperbolic systems of equations.Communications on Pure and Applied Mathematics, 18:697–715,

    James Glimm. Solutions in the large for nonlinear hyperbolic systems of equations.Communications on Pure and Applied Mathematics, 18:697–715,

    work page

  5. [4]

    A generalized riemann problem for quasi-one-dimensional gas flows.Advances in Applied Mathe- matics, 5(1):1–30, 1984

    James Glimm, George Marshall, and Bernd Plohr. A generalized riemann problem for quasi-one-dimensional gas flows.Advances in Applied Mathe- matics, 5(1):1–30, 1984. MR0730835, Zbl 0532.76071

    work page arXiv 1984

  6. [5]

    Jagtap, George Em Karniadakis, and Kenji Kawaguchi

    Zheyuan Hu, Ameya D. Jagtap, George Em Karniadakis, and Kenji Kawaguchi. Augmented physics-informed neural networks (apinns): A gat- ing network-based soft domain decomposition methodology.Engineering Applications of Artificial Intelligence, 123:107183, 2023

    work page 2023

  7. [6]

    Jagtap and George Em Karniadakis

    Ameya D. Jagtap and George Em Karniadakis. Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations. Communications in Computational Physics, 28(5):2002–2041, 2020

    work page 2002

  8. [7]

    Jagtap, Ehsan Kharazmi, and George Em Karniadakis

    Ameya D. Jagtap, Ehsan Kharazmi, and George Em Karniadakis. Con- servative physics-informed neural networks on discrete domains for conser- vation laws: Applications to forward and inverse problems.Computers & Mathematics with Applications, 109:147–166, 2021

    work page 2021

Show all 13 references
  1. [9]

    LeFloch and Shuyang Xiang

    Philippe G. LeFloch and Shuyang Xiang. Weakly regular fluid flows with bounded variation on the domain of outer communication of a schwarzschild black hole spacetime. ii.Journal de Math´ ematiques Pures et Appliqu´ ees, 122:272–317, 2019

    2019

  2. [10]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics- informed neural networks: A deep learning framework for solving for- ward and inverse problems involving nonlinear partial differential equa- tions.Journal of Computational Physics, 378:686–707, 2019. 10

    2019

  3. [11]

    Central schemes for conservation laws with application to shallow water equations

    Giovanni Russo. Central schemes for conservation laws with application to shallow water equations. In S. Rionero and G. Romano, editors,Trends and Applications of Mathematics to Mechanics: STAMM 2002, pages 225–246. Springer, 2005. MR2162629, Zbl 1077.76046

    2002

  4. [12]

    Adaptive mesh refinement for hyperbolic systems based on third-order compact weno reconstruction.Journal of Scientific Computing, 66(2):692–724, 2016

    Michele Semplice, Andrea Coco, and Giovanni Russo. Adaptive mesh refinement for hyperbolic systems based on third-order compact weno reconstruction.Journal of Scientific Computing, 66(2):692–724, 2016. MR3449835, Zbl 1329.65067. 11

    2016

  5. [1965]

    MR0194770, Zbl 0142.39602

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.