{"id":"d98c0456-0804-450d-82fe-a8496c2416be","arxiv_id":"2506.12922","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Standard physics-informed neural networks reproduce analytic solutions for five coupled Burgers benchmarks with small reported errors, but the paper omits code, training details, and architectures needed to verify or reuse them.","lead":"The paper trains physics-informed neural networks to solve five Burgers equation test problems and reports small errors versus exact solutions. It claims the networks beat several classical solvers, but gives no code, no training sizes, and no selected network architecture, so the runs cannot be repeated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The only explicitly stated PINN residual and initial condition in Section 4 do not match Example 2, the source of the headline superiority comparison, so Tables 5.2 and 5.3 cannot be reproduced from the manuscript as written.","rationale":"The reader correctly identifies the lack of architecture and training details as a reproducibility failure, and the conclusion of the paper cannot be checked from the manuscript. My stress-test pass found a more specific, internally checkable problem: Section 4, which is the only place where the loss function is written down, gives residuals and an initial condition that match Example 3 but not Example 2, the example used for the headline comparison with prior methods. For Example 2, the Section 4 residual differs from the stated equation in both the u*u_x term (sign) and the (u*v)_x coefficient, so the printed residual is nonzero at the exact solution. This means that either the text does not describe the actual experiment or the experiment used a different residual; in either case Tables 5.2 and 5.3 cannot be regenerated from the paper. I am not claiming the numerical results are false; PINNs for Burgers-type equations are plausible and the order of magnitude of the reported errors is not surprising. The issue is that the paper as written is internally inconsistent and underspecified, so no referee can determine what was computed. A concrete residual check, followed by a reproduction of Example 2 with a disclosed configuration, would settle whether the claimed superiority is real. Until then the central claim should be treated as unverdictable rather than conditionally acceptable.","tokens_in":9679,"tokens_out":14014,"duration_ms":147852,"concrete_test":"Take the exact solution of Example 2 (u=0.05*(1 - tanh(0.00625*(x+0.0125*t))), v=0.05*(-0.5 - tanh(0.00625*(x+0.0125*t)))) and evaluate the Section 4 residual f_u = u_t - u_xx + 2*u*u_x - (u*v)_x on a grid at t=0.5. If the maximum absolute residual is not near machine precision but is of order 1e-5 or larger, the printed loss is not consistent with the stated PDE of Example 2. Then re-run the Example 2 experiment using the residual actually implied by Eq. (5.2) and report the resulting Table 5.2 errors; if those errors differ materially from the published values, the published algorithm is not the algorithm that produced the table.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central quantitative claim is that PINN gives better and more effective results than [24,35,3,1], based on the small errors in Tables 5.2 and 5.3 for Example 2. But the algorithm specification in Section 4 is not the specification needed for that example. Section 4 defines the residual as f_u = u_t - epsilon*u_xx + 2*u*u_x - (u*v)_x and the initial condition as u(x,0)=v(x,0)=cos(pi*x). Example 2 in Section 5 instead solves u_t - u_xx - 2*u*u_x + 0.1*(u*v)_x = 0 for x in (-10,10), with tanh-type initial profiles. Substituting the stated exact solution of Example 2 into the Section 4 residual gives a nonzero value (about -2.7e-5 at t=0.5, x=0), so the exact solution is not a zero-residual point of the printed loss. The paper never states which residual was actually used for Examples 1, 2, 4, or 5; only Example 3 matches the Section 4 formulas. No code, no grid, and no training configuration are supplied, and the Data Availability statement explicitly says none will be made available. Consequently, even the formal content of the paper does not determine the algorithm that generated Tables 5.2 and 5.3, and the claimed superiority over Bak, Ahmad, Khater, and Rashid is not verifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed neural network (PINN) framework for simulating nonlinear Burgers-type equations, and demonstrates it on five test problems: 1D coupled, 2D single, and 2D coupled Burgers models. The reported results, especially for Example 2, are compared against existing methods (Bak, Ahmad, Khater, Rashid) in terms of L2 and L-infinity errors, and the conclusion claims that PINN offers better and more effective results than those references. The manuscript also describes the network architecture, loss function, and training procedure, though without providing the numerical details needed for reproduction.","tokens_in":10003,"tokens_out":7295,"duration_ms":73442,"significance":"If the reported results were reproducible and the comparisons fair, the paper would provide a useful benchmark for applying PINNs to coupled Burgers-type equations, particularly in the high-Reynolds-number regime and in 2D. The exact solutions quoted for Examples 1–3 appear to satisfy the stated equations, which is a strength. However, the manuscript as written does not establish its central quantitative claims: the printed loss function and initial condition in Section 4 match only Example 3, not the Example 2 source of the headline comparison; the training configuration is under-specified; and the reported tables do not uniformly support the claimed superiority over the cited baselines. These issues undermine the verifiability of the error tables and the validity of conclusion (iii).","major_comments":[{"comment":"Section 4 defines the PDE residuals as f_u = u_t − εu_xx + 2u u_x − (uv)_x and the initial condition as u(x,0)=v(x,0)=cos(πx). This matches only Example 3, whose exact solution is e^{−ε π² t} cos(πx). Example 2, the only example with a comparison to prior methods, uses u_t − u_xx − 2u u_x + 0.1(uv)_x = 0 with tanh-type initial profiles. Substituting the Example 2 exact solution into the Section 4 residual gives a nonzero value (approximately −2.7e−5 at t=0.5, x=0), so the printed loss cannot be the loss that produced Tables 5.2 and 5.3. The paper never states the residual actually minimized for Examples 1, 2, 4, or 5.","section":"§4 vs §5, Example 2"},{"comment":"The manuscript does not provide the information needed to reproduce the numerical results. Section 3.1 reports a hyperparameter sweep over L ∈ {3,4,5,6,7} and H ∈ {20,30,40,50,60} but never states the selected depth and width. Algorithm Step 4 says 'For E epochs' but E is never defined. The paper does not report the number of collocation, initial-condition, or boundary-condition points, the random seed, or any learning-rate schedule beyond the generic Adam defaults. The Data Availability statement says no data will be made available, and no code is referenced. Consequently, the error values in Tables 5.1–5.3 and the figures cannot be independently verified.","section":"§3.1, §5 Algorithm step 4"},{"comment":"The claim that 'PINN demonstrates significantly lower L∞ and L2 errors for both u and v compared to existing methods' is contradicted by the tables themselves. At t=0.5, Ahmad et al. [1] reports L∞ = 2.1840e−4 for u and 2.5169e−4 for v, while the PINN errors are 3.2962e−4 and 6.4322e−4, respectively. Thus the PINN is worse than at least one baseline at the earliest reported time. In addition, Table 5.3 labels the PINN columns as 'error_u' although the table is for v, and several entries for Khater and Rashid are missing. The comparison also gives no information about the grids, norms, or stopping criteria used in the baseline papers, so the conclusion of superiority is not established.","section":"Tables 5.2 and 5.3, §6 (iii)"},{"comment":"The abstract and Section 6 state that the method is validated through L2 and L-infinity error norms, but Examples 4 and 5 (2D single and 2D coupled Burgers) are presented only through figures (Figs. 5.6–5.8), with no error tables or convergence measurements. Without quantitative error metrics for these two examples, the claim of rigorous validation for higher-dimensional cases is not supported.","section":"§5, Examples 4 and 5"}],"minor_comments":[{"comment":"The phrase 'competitive performance in terms of inaccuracy' should be 'accuracy'; 'inaccuracy' is presumably a typo.","section":"Abstract"},{"comment":"The text says the method is applied to 'three different example' but the paper presents five examples; the count should be corrected.","section":"§5, opening paragraph"},{"comment":"The sentence 'Errors shown in Table 5.2 of the problem are an order of magnitude smaller than in problems 1–2' should refer to 'Example 1', since this is Example 2.","section":"§5, Example 2 discussion"},{"comment":"The symbol ε is used both for the PDE viscosity coefficient and for the Adam optimizer's small constant in Eq. (4.2); this notational collision could confuse readers.","section":"Eq. (4.2)"},{"comment":"The notation u·∇u is ambiguous in the 1D setting where u is a scalar; the examples later use u u_x, so the general formulation should clarify the intended interpretation.","section":"§2, Eq. (2.3)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim of superiority over earlier methods is not verifiable from the manuscript because the loss specification and training details do not match the reported experiments. This is fixable by adding per-example residual definitions, full hyperparameter and sampling details, and a corrected comparison. The contribution is otherwise incremental, being a standard PINN application to Burgers-type equations, but a thorough revision addressing reproducibility and the overstatement in the conclusion could make it acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a straightforward PINN application to a spread of smooth Burgers benchmarks. There is no new method — the framework is Raissi et al. [34] — and the useful part is the breadth of examples: 1D coupled, 2D single, 2D coupled, and a high-Re (10^6) case. The quoted analytic solutions for Examples 1–3 do satisfy the PDEs as written, so the benchmark setup is not fabricated.\n\nWhat the paper does well: the overview sections are clear, and the authors are honest that they are using the standard tanh/Adam setup. The error tables alone look plausible for these smooth solutions, and the high-Reynolds test is a reasonable stress case even though the exact solution is very smooth.\n\nNow the soft spots, and they are serious. No code, no data, and the Data Availability statement says none will be provided. The hyperparameter sweep is described but the selected depth and width are never reported; neither are collocation point counts, epoch counts, or random seeds. More damning, Section 4's residual and initial condition do not match Examples 1, 2, 4, or 5 — they match only Example 3. Example 2, which is the basis of the claimed superiority over Bak, Ahmad, Khater, and Rashid in Tables 5.2 and 5.3, has different coefficients (0.1 and 0.3 multiplying (uv)_x) and tanh initial profiles. Substituting the Example 2 exact solution into the Section 4 residual gives a nonzero value, so the printed loss is not the loss that generated those tables. The comparisons with prior methods are also uncontrolled: there is no evidence the earlier methods were evaluated on the same grids, norms, or stopping criteria.\n\nThe central claim — that PINNs can approximate these smooth Burgers solutions — is almost certainly true and already established in the literature, so the conclusion is not dangerously wrong. But the evidence as presented is unverifiable. A revision that ships code, states the selected architecture and training configuration, and fixes the Section 4 mismatch would turn this into a modest but useful data point for practitioners.\n\nThis paper is for readers who want another confirmation that standard PINNs work on smooth Burgers problems and who are not bothered by closed reproducibility. My recommendation: do not send to peer review in its current form. The internal mismatch and missing code mean a referee would have to take the headline numbers on faith. If the authors provide the missing artifacts and correct the inconsistency, it becomes a publishable contribution. As is, desk reject with an invitation to resubmit with code and fixes.","headline":"Standard PINN applied to smooth Burgers benchmarks, but the headline accuracy comparisons are unreproducible because the printed residual does not match Example 2 and no code, training details, or selected architecture are given.","tokens_in":10552,"tokens_out":3133,"would_cite":false,"duration_ms":32697,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N15","65N30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a physics-informed neural network trained on PDE residuals, initial data, and boundary data accurately simulates 1D and 2D nonlinear Burgers-type models and, on the paper's comparison tables, beats four classical…","keywords":["physics-informed neural networks","Burgers' equation","coupled nonlinear PDEs","mesh-free methods","automatic differentiation","L2 error","L-infinity error","Reynolds number"],"falsifier":"Re-running the paper's Example 2 with the stated loss weights and Adam settings while recording the actual architecture size and training point counts would settle whether the printed $L^\\infty$ and $L^2$ errors for both solution components reproduce across random seeds; if the rerun error varies widely or fails to stay below the four baseline values, the claimed superiority would not hold up.","tokens_in":9461,"feed_emoji":"🌊","tokens_out":11704,"duration_ms":106935,"temperature":0.7,"pith_summary":"The paper is trying to establish that a physics-informed neural network (PINN), a neural network trained to minimize the governing equation's residual together with initial and boundary conditions, can act as a mesh-free solver for nonlinear Burgers-type models. It runs five test problems covering a 1D coupled system, a modified coupled system, a high-Reynolds-number regime, a 2D scalar equation, and a 2D coupled system, and reports $L^2$ and $L^\\infty$ errors against exact solutions. The authors claim the PINN results are competitive with, and on the modified coupled system better than, the four comparison methods cited in their tables. If the claim is right, nonlinear time-dependent PDEs can be solved without building a mesh, which matters for high-dimensional or irregular domains where classical discretization becomes expensive.","feed_headline":"Mesh-free PINNs beat four classical solvers on Burgers systems","feed_subtitle":"The method's reported error norms on 1D and 2D nonlinear Burgers-type models undercut the cited baselines.","key_machinery":"The load-bearing object is the composite physics-informed loss $L_{\\mathrm{total}} = L_{\\mathrm{PDE}} + 10 L_{\\mathrm{IC}} + 10 L_{\\mathrm{BC}}$, where $L_{\\mathrm{PDE}}$ is the mean squared residual of the Burgers-type system evaluated with automatic differentiation at collocation points, and $L_{\\mathrm{IC}}$ and $L_{\\mathrm{BC}}$ are mean squared mismatches of initial and boundary conditions. The trial solution is a fully connected feed-forward network with tanh activations that maps space-time coordinates to the solution components; because derivatives come from automatic differentiation, no mesh or discretized grid appears in the algorithm. The paper sweeps depth $L \\in \\{3,4,5,6,7\\}$ and width $H \\in \\{20,30,40,50,60\\}$ to select an architecture, then trains with the Adam optimizer at learning rate $10^{-3}$.","core_discovery":"The central claim is that a fully connected feed-forward network with tanh activations, taking space-time coordinates as input and returning the solution components as output, can approximate the exact solutions of the coupled Burgers systems when trained to minimize the composite loss $L_{\\mathrm{total}} = L_{\\mathrm{PDE}} + 10 L_{\\mathrm{IC}} + 10 L_{\\mathrm{BC}}$. The PDE residual is evaluated by automatic differentiation, so no spatial discretization appears in the method. The paper's strongest quantitative assertion is in its second example, where the reported $L^\\infty$ and $L^2$ errors for the first solution component are lower at the tabulated times than the errors it attributes to the four cited baseline methods; the conclusion states that PINN offers better and more effective results than those references. The paper also reports stable solutions at Reynolds number $10^6$, a regime where conventional mesh-based schemes often require stabilization.","pith_inferences":["Beyond the paper, the same composite loss could be applied to other coupled parabolic systems such as reaction-diffusion or shallow-water equations, since the machinery is problem-agnostic apart from the residual definition.","Beyond the paper, the fixed weighting of the initial and boundary loss terms is a knob worth tuning adaptively; loss reweighting could reduce the boundary-adjacent discrepancies the paper itself observes in its first example.","Beyond the paper, the claimed advantage over the cited baselines could be tested directly by re-implementing those methods on identical grids and norms instead of importing published error numbers."],"forward_implications":["If the central claim is correct, nonlinear coupled Burgers problems can be simulated without generating a mesh, which lowers the setup cost for problems on irregular or high-dimensional domains.","The reported stability at Reynolds number $10^6$ suggests the same network formulation could handle advection-dominated flows where classical schemes need added stabilization.","Since one architecture and loss recipe covers 1D coupled, 2D scalar, and 2D coupled models, the result implies the method transfers across PDE systems with minimal reformulation.","If the comparison tables are accepted, PINNs become a practical default benchmark solver for nonlinear Burgers-type equations and similar convection-diffusion systems.","Because the trial solution is fully differentiable, the trained representation can supply gradients needed for downstream sensitivity analysis or optimization without extra discretization."],"supporting_citations":[{"why":"This reference supplies the foundational PINN method: train a network by minimizing the PDE residual together with initial and boundary losses.","marker":"[34]"},{"why":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat.","marker":"[3]"},{"why":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat.","marker":"[1]"},{"why":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat.","marker":"[24]"},{"why":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat.","marker":"[35]"},{"why":"This reference supplies the Adam optimizer whose update rule and default settings are used throughout training.","marker":"[25]"},{"why":"This reference supplies the exact solution used for the high-Reynolds-number example where the paper reports stable behavior.","marker":"[40]"}],"fun_headline_variants":["PINN beats four solvers on nonlinear Burgers equations","Mesh-free PINN tames Burgers systems at Re=1e6","Autodiff PINN outperforms classical solvers on Burgers","Physics-informed networks solve coupled Burgers without grids"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The superiority claim rests on the error tables being reproducible from the setup as described and on the cited comparison errors being computed under equivalent conditions, but the paper does not report the chosen depth and width, the number of collocation, initial, and boundary points, the number of training epochs, the random seed, or the stopping criterion.","fun_headline_variants_meta":{"raw":{"variants":["PINN beats four solvers on nonlinear Burgers equations","Mesh-free PINN tames Burgers systems at Re=1e6","Autodiff PINN outperforms classical solvers on Burgers","Physics-informed networks solve coupled Burgers without grids"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2912,"prompt_tokens":901,"completion_tokens":2011,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":1940}},"tokens_in":517,"tokens_out":2011,"duration_ms":16613,"temperature":1.0,"reasoning_tokens":1940,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:37:27.302629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the paper's Example 2 with the stated loss weights and Adam settings while recording the actual architecture size and training point counts would settle whether the printed $L^\\infty$ and $L^2$ errors for both solution components reproduce across random seeds; if the rerun error varies widely or fails to stay below the four baseline values, the claimed superiority would not hold up.","supporting_citations":[{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"This reference supplies the foundational PINN method: train a network by minimizing the PDE residual together with initial and boundary losses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat."},{"cited_title":"Ahmad, T","cited_arxiv_id":null,"evidence_quote":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat."},{"cited_title":"Khater, R","cited_arxiv_id":null,"evidence_quote":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat."},{"cited_title":"Rashid and A","cited_arxiv_id":null,"evidence_quote":"This reference provides one of the four comparison baselines in Example 2 that the paper claims to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This reference supplies the exact solution used for the high-Reynolds-number example where the paper reports stable behavior."}],"review_version":1}