{"id":"6244dfef-8074-44ce-8f0d-2d5416e4be55","arxiv_id":"2506.00951","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A shock-aware physics-informed neural network can reproduce smooth and discontinuous solutions of the relativistic Burgers equation in a Schwarzschild spacetime, though without quantitative validation.","lead":"This paper trains a physics-informed neural network to solve a relativistic Burgers equation outside a Schwarzschild black hole, adding a shock-aware block to represent discontinuities. It reports that the network can approximate smooth and shock solutions, but with lower accuracy than classical numerical methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Godunov flux formula in §3.2 is incorrect for Riemann data with v_L > 0 > v_R and negative shock speed, so the PINN residual enforces the wrong conservation law and the central robustness claim is unsupported.","rationale":"The reader's weakest-assumption analysis correctly identifies the most load-bearing defect: the Godunov flux formula in §3.2, which is the core of the residual construction, is wrong for a natural class of Riemann data. For v_L > 0 > v_R with negative shock speed, the formula returns min{f(v_L), f(v_R)} = f(v_L) instead of the exact Godunov flux f(v_R). Since the loss function minimizes this residual, training would enforce a different numerical flux and hence a different discrete conservation law in that regime. The paper's claimed contribution is a robust PINN for shocks in curved spacetime, so this is not a peripheral typo; it directly undermines the generality of the central claim. The paper also provides no quantitative error metrics or reference comparisons, and Section 5 concedes accuracy is far below traditional methods, but the flux error alone is sufficient to block acceptance. I do not find a separate, more severe concern that would change the reader's verdict; the other weaknesses (missing boundary condition details, absence of quantitative validation) reinforce but do not replace this one. I am not treating the paper's unconventional presentation as evidence of bad faith; the mathematical issue stands on its own. A single targeted computation and re-training on the untested branch would settle whether the flaw is merely theoretical or manifests as incorrect shock dynamics. Because the reader's rejection is already based on this concern, my verdict recommendation is unchanged.","tokens_in":5616,"tokens_out":5429,"duration_ms":55462,"concrete_test":"Evaluate the §3.2 flux formula for v_L = 0.2, v_R = -0.5, metric = 1 (so s < 0): the formula returns min{f(0.2), f(-0.5)} = f(0.2) = -0.48, whereas the exact Godunov flux is f(v_R) = f(-0.5) = -0.375. Then run the Section 4 training procedure on Riemann data with these left and right states; compare the learned shock speed with the exact entropy solution speed s = (v_L + v_R) metric(r) / 2. If the learned shock follows the incorrect left-state flux, the concern is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed PINN robustly handles shock discontinuities for the relativistic Burgers equation on a Schwarzschild exterior. That claim depends on the residual in §3.2 being a correct discrete form of (1.1). The Godunov flux formula is not correct for all Riemann data. For the branch v_L > v_R, s < 0, the formula returns min{f(v_L), f(v_R)}. For convex f(v) = (v^2 - 1)/(2(1 - 2M/r)), when v_L > 0 > v_R and v_L + v_R < 0 (so s = (v_L + v_R) metric(r)/2 < 0), one has |v_R| > |v_L| and therefore f(v_R) > f(v_L); hence min{f(v_L), f(v_R)} = f(v_L). The exact Godunov flux for a left-moving shock is f(v_R), the right state. Thus the residual in this branch is anti-upwind and would drive the network toward a different effective conservation law. The Section 4 experiments never exercise this branch: the steady-shock test has s = 0, and the moving-shock test uses states of the same sign, so the flaw is hidden. Since the paper's stated contribution is general shock capability, this is a load-bearing mathematical error rather than a mere implementation detail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":5968,"tokens_out":8189,"duration_ms":75738,"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":[{"comment":"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.","section":"Section 3.2"},{"comment":"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.","section":"Section 4"},{"comment":"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.","section":"Sections 3.2–3.3"}],"minor_comments":[{"comment":"The phrase 'pros and cos' should be 'pros and cons'.","section":"Section 5"},{"comment":"The second branch of the piecewise definition says 'r0 > 0', which should likely be 'r > r0'.","section":"Section 2.1, Eq. (2.4)"},{"comment":"The notation 'σ net(t)' should be 'σ_net(t)' for readability.","section":"Section 3.1"},{"comment":"The word 'Rankie-Hugoniot' is a typo for 'Rankine–Hugoniot'.","section":"Section 3.2"},{"comment":"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.","section":"Section 2.1, Theorem 2.1"}],"recommendation":"reject","confidential_remarks":"The paper is very short and leans heavily on the author's own prior work [8,9] for the physical theory. The Godunov flux error, while not affecting the presented tests, is a fundamental flaw in the proposed method and must be corrected before any resubmission. The absence of quantitative validation is another serious gap. The manuscript would need substantial additional work with new experiments, error analysis, and reproducibility details to be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The architecture is the real news: a shock-aware jump block with a learned shock location, combined with a Godunov-inspired residual, applied to the relativistic Burgers equation on Schwarzschild. That combination is new and sensible, and the three numerical examples do show the method can capture smooth and shock solutions in this setting. The paper also deserves credit for explicitly admitting that the accuracy is far below traditional methods like Glimm's scheme. It does not oversell the practical value.\n\nThe soft spot is a genuine mathematical error in the Godunov flux formula in Section 3.2. For the branch v_L > v_R with negative shock speed, the exact Godunov flux is f(v_R), but the formula returns min{f(v_L), f(v_R)}, which is anti-upwind when |v_R| > |v_L|. That is not a cosmetic slip; it changes the conservation law the network is trained to satisfy. The paper's experiments never exercise this branch: the steady shock has s = 0, and the moving shock uses states of the same sign, so the error is hidden. Since the stated contribution is a robust PINN for shocks in curved spacetime, this is load-bearing. A method that is wrong on an untested but generic set of Riemann data cannot support that claim.\n\nThe paper also lacks quantitative validation: no error norms, no reference comparisons, and the boundary data are not specified. The loss weights and the boundary conditions are parameters but their values are not given. These are serious gaps, but they are fixable. The honest limitation statement in Section 5 is a point in the paper's favor.\n\nWho is this for? Researchers interested in PINN architectures for conservation laws on curved metrics, especially those who want a mesh-free solver with low accuracy tolerance. It is also a useful cautionary example of how a flawed numerical flux can hide inside a learned residual.\n\nMy recommendation: send it to peer review. The architectural idea is legitimate and the problem is relevant, but the manuscript needs major revision to fix the flux formula, add quantitative comparisons, and restrict or rephrase the robustness claim. A good referee would catch this quickly. I would not cite it in its current form.","headline":"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.","tokens_in":6428,"tokens_out":2913,"would_cite":false,"duration_ms":29235,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M70","68T07","35L65","83C57"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["physics-informed neural networks","relativistic Burgers equation","Schwarzschild spacetime","shock waves","Godunov flux","conservation laws","mesh-free methods","general relativity"],"falsifier":"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.","tokens_in":5416,"feed_emoji":"🕳️","tokens_out":8146,"duration_ms":67941,"temperature":0.7,"pith_summary":"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.","feed_headline":"Neural network reproduces shock waves near a black hole","feed_subtitle":"A physics-informed net with a shock-location block handles smooth and discontinuous flows on Schwarzschild spacetime.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the Glimm-method numerical solutions for the relativistic Burgers equation on a Schwarzschild exterior that the paper uses for qualitative comparison and to position its accuracy.","marker":"[8]"},{"why":"Gives the existence theory and the explicit steady-state and shock solutions that the paper uses to construct initial conditions and to check whether the learned solutions match the expected weak solutions.","marker":"[9]"},{"why":"Establishes the standard PINN framework and loss construction that the paper extends with the shock-aware block and Godunov-inspired residual.","marker":"[10]"},{"why":"Introduces conservative PINNs for conservation laws, which motivates the paper's use of a numerical flux in the residual to respect jump conditions.","marker":"[7]"}],"fun_headline_variants":["Neural network simulates shock waves near a black hole","PINN solves relativistic Burgers equation in black hole exterior","Physics-informed net handles shocks in curved spacetime","AI predicts discontinuities in Schwarzschild geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Neural network simulates shock waves near a black hole","PINN solves relativistic Burgers equation in black hole exterior","Physics-informed net handles shocks in curved spacetime","AI predicts discontinuities in Schwarzschild geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000606,"raw_usage":{"total_tokens":2774,"prompt_tokens":841,"completion_tokens":1933,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":1872}},"tokens_in":457,"tokens_out":1933,"duration_ms":13654,"temperature":1.0,"reasoning_tokens":1872,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:54:01.654697+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"LeFloch and Shuyang Xiang","cited_arxiv_id":null,"evidence_quote":"Provides the Glimm-method numerical solutions for the relativistic Burgers equation on a Schwarzschild exterior that the paper uses for qualitative comparison and to position its accuracy."},{"cited_title":"LeFloch and Shuyang Xiang","cited_arxiv_id":null,"evidence_quote":"Gives the existence theory and the explicit steady-state and shock solutions that the paper uses to construct initial conditions and to check whether the learned solutions match the expected weak solutions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the standard PINN framework and loss construction that the paper extends with the shock-aware block and Godunov-inspired residual."},{"cited_title":"Jagtap, Ehsan Kharazmi, and George Em Karniadakis","cited_arxiv_id":null,"evidence_quote":"Introduces conservative PINNs for conservation laws, which motivates the paper's use of a numerical flux in the residual to respect jump conditions."}],"review_version":1}