{"id":"711d6fd8-c6f4-4e0c-a0a1-e6912915b839","arxiv_id":"2501.11372","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PINNs with a reweighted variable-conservation form solve 2D shallow water cases with rainfall and terrain, but the claimed theoretical superiority is only partially supported.","lead":"The authors test whether physics-informed neural networks can simulate two-dimensional flood flow over terrain with rainfall, using eleven benchmark cases and open-source code. They find the networks can match classical solvers closely, and that a hybrid equation form trains better than alternatives, while an energy-based condition hurts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 2.3's proof of variable-conservation superiority rests on a false inference, and dDB8 never compares against the conservative form; the central form-selection claim is not established.","rationale":"I chose this as the most load-bearing concern because it attacks the paper's central methodological contribution, not just its reproducibility. The abstract and conclusion state that theoretical analysis and experimental validation demonstrate the variable-conservation form is superior to both alternatives. The proof's key inference is plainly invalid, and the single supporting experiment omits the very comparator, the conservative form, against which the claim is made. Even if the dimensional transformation and per-case weights were fully specified, the form-selection claim would remain unproven. The feasibility part of the paper, including static well-balance, rainfall cases, dynamic dam break, and tidal problems, is credible as a set of demonstrations, and the open-source module is valuable, so the reader's CONDITIONAL verdict remains appropriate. My read does not change that verdict; it adds a specific condition: either fix the proof or run the conservative-form comparison before claiming superiority over it. I partly agree with the reader's weakest assumption; the undefined transformation is a real reproducibility concern, but I find the invalid Section 2.3 argument and the missing conservative-form experiment more damaging to the central claim.","tokens_in":20264,"tokens_out":9477,"duration_ms":93692,"concrete_test":"Run dDB8 with the conservative form (Eq. 5, network outputs h, hu, hv) under the exact dDB8 recipe: 4x300 tanh, N-LAAF, Adam 1e-3 with 1e-2 decay at step 20,000, 30,000 steps, differential/entropy weights all 1, IC weight 100, BC weight 1, positivity/dry-wet weight 10. Evaluate h, u, v MAE/RMSE on the same test grid used for Figure 3. If conservative-form errors are comparable to or better than PINN_vc's, the central claim that variable-conservation is more reliable than the conservative form is unsupported by the only form-comparison experiment; if they are substantially worse, the empirical half survives, but the Section 2.3 proof must still be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim includes that the hybrid variable-conservation form is a more reliable choice than the conservative or variables form. The theoretical support in Section 2.3 is unsound. From the primitive-form optimality condition (nabla_{theta*} r)^T r = 0, the paper writes: 'Considering that r is not necessarily 0, this implies nabla_{theta*} r = 0.' This is false: (nabla r)^T r = 0 is a sum of three scalar products and can hold with r nonzero and nabla r nonzero. Because this false step is used to argue that theta* is not a minimizer of the variable-conservation loss and that the two forms have different local minima, the theoretical conclusion that the variable-conservation form is 'less prone to getting trapped in local minima' does not follow. The only experiment behind the form comparison, dDB8, compares the primitive form, the variable-conservation form, and the variable-conservation form with entropy (Figure 3); the conservative form named in the abstract and in Section 2.3 is never run. Thus the head-to-head superiority claim is supported neither by the flawed proof nor by a direct conservative-form comparison. The undefined 'dimensional transformation' is a separate reproducibility gap; my concern is that the claimed result itself is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates the use of physics-informed neural networks (PINNs) for solving the two-dimensional shallow water equations (SWE) with terrain topography and rainfall source terms. The authors combine PINNs with a dimensional transformation technique and N-LAAF, and they introduce several loss components for non-negative water depth, dry-wet conditions, and an entropy stability condition. The manuscript reports eleven case studies: three static water cases on bump, depression, and tidal topographies; four static cases with rainfall; and four dynamic cases (pseudo-2D dam break, circular dam break, tidal flow, and tidal flow with rainfall). Errors are measured against exact solutions or high-resolution numerical references (HLL and entropy-stable schemes). The paper additionally argues, through a gradient-based theoretical analysis in Section 2.3 and the pseudo-2D dam break experiment, that a hybrid variable-conservation form of the SWE is preferable to the variables form and the conservative form. It also reports that adding an entropy condition does not improve accuracy and can cause training failure. An open-source module on the PINNacle platform is released.","tokens_in":20528,"tokens_out":9443,"duration_ms":84986,"significance":"If the central feasibility claim holds, this paper is a useful contribution to the growing literature on PINNs for 2D free-surface flows, especially because it includes rainfall and complex topography, which are absent from most prior PINN-SWE studies. Strengths include the breadth of benchmarks (eleven cases), the use of external numerical references (exact Riemann solution, HLL, ES1) rather than self-comparison, the consistently small reported MAE/RMSE in the static cases, and the open-source code with a data availability statement. However, the paper's headline claim of form superiority is not currently supported: the theoretical argument in Section 2.3 contains a logical error, and the only experimental form comparison (dDB8) omits the conservative form entirely. The undefined 'dimensional transformation technique' also prevents replication of the successful static rainfall cases. The conclusion about the entropy condition is interesting but rests on a limited set of experiments. Overall, the feasibility result is credible, but the main comparative and theoretical claims require substantial revision before the paper can be accepted.","major_comments":[{"comment":"The inference \"Considering that r is not necessarily 0, this implies ∇θ*r = 0\" is incorrect: the necessary condition (∇θ*r)^T r = 0 is a sum of three scalar products and does not force the Jacobian ∇θ*r to vanish. Consequently, the deduction that θ* is not a minimizer of the variable-conservation loss (Eq. 42) does not follow. Even if the two loss functions had different local minima, that would not imply that the variable-conservation form is 'less prone to getting trapped in local minima' without further properties of the loss landscape. Thus the theoretical support for the superiority claim in the abstract and in Section 2.3 is not established.","section":"Section 2.3, Eqs. (40)-(42)"},{"comment":"The experimental comparison of equation forms in case dDB8 includes only the primitive variables form, the variable-conservation form, and the variable-conservation form with entropy stability. The conservative form is never run, despite being named in the abstract and Conclusion as one of the two alternatives. The only evidence against the conservative form is the unreported 'preliminary experiments' mentioned in Section 2.3. To support the claim that the variable-conservation form is superior to both the variables form and the conservative form, the authors should either run the conservative form in the same benchmark or substantially weaken the claim.","section":"Section 4.3, Figure 3"},{"comment":"The 'dimensional transformation technique' is stated to be used in cases sFR4, sBR5, sDR6, and sTR7 (Section 3.2) and is advertised in the abstract as part of the method, but its definition is never given. Without specifying the transformation of coordinates or variables and the inverse transformation, the experiments cannot be reproduced and the claimed benefit of this technique cannot be assessed. Please provide the exact transformation formulas and state which cases employ them.","section":"Section 3.2 and Section 3.4"}],"minor_comments":[{"comment":"The caption lists the four static rainfall cases as 'sFR5, sBR5, sDR6, and sTR7', but the row label in the table and Section 3.2 use sFR4. Please correct the case numbering for consistency.","section":"Table 2 caption"},{"comment":"The reference for PINNacle (Zhu et al., 2023) has a garbled author list ('Zhongkai Hao Zhu' and a lone 'Jun'). Please correct the author names and citation format.","section":"References"},{"comment":"The subsection uses F(uθ) for the residual after defining r and R; please clarify the relationship between F, r, and R to avoid confusion.","section":"Section 2.3, notation"},{"comment":"The rainfall source term in the entropy inequality is written as g(h+b)p, but b is not defined anywhere. It should presumably be the topography z (or b should be defined explicitly).","section":"Equation (24)"},{"comment":"The loss weights are case-specific (e.g., 1,10,10,1 for static cases; 10,1,1 for circular dam break and tidal cases; 1,1,1 for dDB8), but no sensitivity analysis is provided. A brief discussion of robustness to these choices would strengthen the reproducibility of the approach.","section":"Section 3.4"},{"comment":"The statement that the variable-conservation form PINN 'generally outperforms' the HLL scheme is not quantified. Reporting the MAE/RMSE of both the PINN and the HLL solution against the exact solution would substantiate this comparison.","section":"Section 4.3, dDB8 discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid open-source codebase and a valuable set of benchmarks for the feasibility claim. The main issues are that the central form-superiority claim is undermined by a logical error in Section 2.3 and by the absence of a conservative-form experiment, and that the dimensional transformation is undefined. These are fixable within a major revision, and I do not see grounds for rejection based on the feasibility results alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a look for the empirical material: eleven PINN cases for 2D shallow water with terrain and rainfall, an open-source module on PINNacle, and two numerical solvers as benchmarks. The feasibility claim—that PINNs can produce low-error solutions on these benchmarks—largely holds up. The static well-balance cases, rainfall cases, and dynamic tidal cases all show small MAE/RMSE, and the volume conservation check in dTR11 is a nice touch. That is the real contribution: a benchmark suite and a demonstration that 2D SWE with source terms is tractable with PINNs.\n\nThe soft spots are real and they land on the paper's central methodological claim, not on the feasibility result. Section 2.3's argument that the variable-conservation form is less prone to local minima rests on a false inference: from (∇_{θ*} r)^T r = 0 they conclude ∇_{θ*} r = 0. That does not follow. The subsequent claim that θ* is not a minimizer of the variable-conservation loss is therefore unsupported. The proof also shows only a reweighting of the primitive-form loss by A^T A, which is a heuristic, not a proof of superiority. On the empirical side, the only head-to-head is dDB8, which compares primitive, variable-conservation, and variable-conservation with entropy—the conservative form named in the abstract is never run. So the form-selection claim is not established.\n\nTwo more reproducibility gaps deserve mention. The 'dimensional transformation' technique is stated to be used in the rainfall cases but never defined. The per-case loss weights are hand-chosen with no sensitivity analysis, and error tables are single runs with no error bars. The accuracy numbers should be read as existence proofs, not robust statistics.\n\nAll that said, the paper is honest about its limitations, ships code and data, and its negative result on entropy regularization is a useful data point. The enthusiasm for PINNs is tempered by actual experiments, which is rare and welcome.\n\nWho is this for? Anyone working on PINNs for hyperbolic conservation laws or flood modeling. It deserves a serious referee—the feasibility result and benchmark suite are valuable, but the form-superiority claim needs either a corrected argument or a downgrade to 'the variable-conservation form worked well in our tests.' I would send it to review with that expectation.\n\nRecommendation: engage with it. Desk rejection would waste a useful empirical study.","headline":"A useful empirical benchmark suite for PINNs on 2D SWE with rainfall and topography, but the claimed theoretical and empirical superiority of the variable-conservation form is not established.","tokens_in":21070,"tokens_out":2196,"would_cite":true,"duration_ms":22391,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that physics-informed neural networks can solve the two-dimensional shallow water equations with terrain and rainfall, and that the hybrid variable-conservation form is a more reliable choice than either the conservative…","keywords":["physics-informed neural networks","shallow water equations","rainfall source terms","terrain topography","well-balanced scheme","variable-conservation form","entropy condition","flood simulation"],"falsifier":"Reproduce case dDB8 (pseudo-2D dam break) with identical architecture, seed, and 30,000 training steps but swap the variable-conservation loss for the primitive-variables loss; if the primitive-variables version matches or beats the hybrid version, the paper's central claim about equation forms would be falsified. A second check: remove the undocumented dimensional transformation from cases sFR4 through sTR7 and keep all other settings unchanged; if accuracy does not degrade, the transformation is not load-bearing.","tokens_in":20005,"feed_emoji":"🌊","tokens_out":11462,"duration_ms":98304,"temperature":0.7,"pith_summary":"Rainstorm floods are usually simulated with finite-volume or finite-difference codes; this paper asks whether physics-informed neural networks (PINNs) can do the same job as an unsupervised, mesh-free alternative. The authors seek to establish that PINNs, helped by a dimensional transformation of the equations and neuron-wise locally adaptive activation functions, can handle two-dimensional shallow water flow over uneven terrain with rainfall source terms. Through eleven test cases, they argue that the shallow water equations should be embedded in a hybrid variable-conservation form, which automatically weights mass and momentum residuals and gives better accuracy and stability than the primitive-variable form. They also report that adding an energy-conservation/entropy penalty does not improve results and can cause training to fail. A sympathetic reader would care because these findings give concrete guidance on how to build PINN flood models and where the method's accuracy currently stands.","feed_headline":"Physics-informed neural networks solve 2D flood equations with rain","feed_subtitle":"A hybrid variable-conservation form beats standard alternatives; entropy penalties can break training.","key_machinery":"The load-bearing object is the residual-coupling matrix $A$ that turns primitive-variable residuals into variable-conservation residuals; the paper shows that the gradient dynamics of the hybrid form descend in characteristic coordinates with weights given by the eigenvalues of $A^T A$, approximately $(1,h^2,h^2)$ at rest, so deeper water automatically receives stronger momentum penalties. This self-weighting is the mechanism claimed to make the hybrid form more stable and less prone to local minima. The other machinery is N-LAAF, the Neuron-wise Local Adaptive Activation Function, which gives each neuron a learnable slope plus a gradient-recovery term, and a dimensional transformation used to rescale variables in the rain cases.","core_discovery":"The paper's central claim is that, for embedding the two-dimensional shallow water equations in a physics-informed neural network, the hybrid variable-conservation form is the right choice. In that form the residuals are $R = A r$, where $r$ are the primitive-variable residuals and $A$ is the $3\\times 3$ matrix with rows $(1,0,0)$, $(u,h,0)$, and $(v,0,h)$; the paper shows that the resulting loss automatically changes into characteristic coordinates and weights each component by the eigenvalues of $A^T A$, approximately $(1,h^2,h^2)$ near static water. Because of this automatic pointwise weighting, the hybrid form is argued to avoid the local minima and instability that plague the primitive-variable form, and in the pseudo-2D dam-break benchmark it matches the exact Riemann solution more closely than a first-order HLL reference. The paper also claims that PINNs with this form, together with a dimensional transformation and N-LAAF activation, solve static water over nontrivial terrain to about $10^{-4}$ m, reproduce rainfall-driven water-level rise to millimeter accuracy, and handle circular dam breaks and tidal flows with and without rain, while adding an entropy-stability penalty gives no real benefit and can over-smooth shocks into training failure.","pith_inferences":["Inference: The matrix-$A$ weighting argument is not limited to shallow water; any hyperbolic system written in quasi-linear form inherits a similar $A$ and could benefit from a hybrid residual form, so the paper's main idea is portable.","Inference: A natural blind test is to apply the same recipe to a new watershed with a different terrain function and a multi-hour rainstorm; the paper's cases run from 0.5 seconds to 5 minutes, so long-horizon accuracy and stability remain open.","Inference: The negative result on the entropy condition is specific to adding it as a soft ReLU-penalized loss; enforcing energy balance through architecture or as a hard constraint might behave differently, so the conclusion should not be over-generalized.","Inference: Since error does not accumulate in time and initial-condition error seems to bound later error, periodic data-assimilation retraining at early times could be a promising route to longer flood forecasts."],"forward_implications":["Static water over bump, depression, and tidal terrain stays well balanced: $h+z$ remains at $0.3$ m and $u,v$ are near zero, with errors between $10^{-5}$ and $10^{-4}$ m.","Rainfall cases show that cumulative water-level rise tracks cumulative rainfall closely, with errors mostly in the $10^{-3}$ to $10^{-5}$ cm range, and the error does not accumulate over time the way time-stepping numerical errors do.","For the pseudo-2D dam break, the variable-conservation form is more accurate than the primitive-variable form and is competitive with a first-order HLL numerical solution, while the primitive form captures wave trends but with large velocity errors.","Adding the entropy condition is not recommended: in static cases it only slightly reduces micro-oscillations, and in the dam-break case it over-smooths the solution and leads to training failure.","Initial-condition accuracy sets the error floor: the approximation error at later times is generally no smaller than the initial-condition error, so high weights on initial conditions and better initial fitting are essential."],"supporting_citations":[{"why":"Defines the physics-informed neural network loss formulation that the paper extends to the 2D shallow water equations.","marker":"[Raissi et al., 2019]"},{"why":"Supplies the N-LAAF locally adaptive activation functions with slope recovery used in all models.","marker":"[Jagtap et al., 2020]"},{"why":"Provides the augmented shallow-water PINN structure, the non-negative depth and dry-wet losses, and the Riemann solver used for the dDB8 benchmark.","marker":"[Dazzi, 2024]"},{"why":"Supplies the neural tangent kernel analysis of loss imbalance that motivates the paper's weighting scheme and large initial-condition weights.","marker":"[Wang et al., 2022]"},{"why":"Defines the entropy-stable, well-balanced ES1 scheme used as the numerical benchmark for the tidal cases.","marker":"[Fjordholm et al., 2011]"},{"why":"Defines the HLL approximate Riemann solver used as reference for circular dam-break and as comparison in the pseudo-2D dam-break case.","marker":"[Toro, 2024d]"},{"why":"Supplies the tidal topography test problem that the paper adapts for static and dynamic tidal cases.","marker":"[Supei et al., 2022]"},{"why":"Provides the benchmark infrastructure on which the paper's open-source shallow-water PINN module is built.","marker":"[Zhu et al., 2023]"}],"fun_headline_variants":["Hybrid PINN form wins for 2D rain floods","Rain terrain floods: hybrid PINN form outperforms","2D flood PINN: hybrid variable-conservation form best","Entropy penalty can break PINN flood training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the dimensional transformation applied to the equations and the per-case loss weights are part of a general recipe rather than tuned to each benchmark; the paper never defines the transformation or gives a rule for choosing the case-specific differential-equation weights, so if those choices were hand-tuned, the reported accuracies and the comparison between equation forms would not transfer automatically.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid PINN form wins for 2D rain floods","Rain terrain floods: hybrid PINN form outperforms","2D flood PINN: hybrid variable-conservation form best","Entropy penalty can break PINN flood training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000628,"raw_usage":{"total_tokens":2947,"prompt_tokens":1034,"completion_tokens":1913,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":1846}},"tokens_in":650,"tokens_out":1913,"duration_ms":16415,"temperature":1.0,"reasoning_tokens":1846,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:20:08.579313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce case dDB8 (pseudo-2D dam break) with identical architecture, seed, and 30,000 training steps but swap the variable-conservation loss for the primitive-variables loss; if the primitive-variables version matches or beats the hybrid version, the paper's central claim about equation forms would be falsified. A second check: remove the undocumented dimensional transformation from cases sFR4 through sTR7 and keep all other settings unchanged; if accuracy does not degrade, the transformation is not load-bearing.","supporting_citations":[],"review_version":1}