{"id":"a65f8a0c-422a-43f2-a36c-e2bd27901c9e","arxiv_id":"2509.02091","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CLINN combines implicit-form, boundedness, and Rankine-Hugoniot loss terms with adaptive refinement, cutting MSE by up to 99.2% on scalar conservation law benchmarks.","lead":"A new neural network solver, CLINN, adds conservation-law constraints (implicit solution, boundedness, shock speed) to the physics-informed loss and adaptively focuses training near shocks. On scalar conservation law benchmarks it reports up to 99.2% lower error than plain PINN.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LRH loss appears to require externally supplied shock speed s(xj,tj); if s comes from the exact Φ given for every benchmark, reported gains are supervised by analytic discontinuity data.","rationale":"The reader's weakest assumption focused on the AN indicator missing double contacts and on the lack of an explicit procedure for constructing Φ. This is closely related but not identical to the most load-bearing issue: the LRH loss term needs the shock speed s(xj,tj), and the paper appears to supply it from analytic Φ in every test case. That would make the comparison to PINN unfair and the method not self-contained, which is more fundamental than indicator accuracy alone. The reader did not explicitly flag the reliance on exact s. I agree with the conditional verdict because the paper could still be salvageable if s can be derived from AN-detected points and if code shows this; however, without that clarification the central empirical claim is not established. I would not escalate to REJECT because the authors may have an undisclosed self-contained procedure; the paper's own text is ambiguous. The proposed test directly checks whether the improvement survives without analytic s. Secondary issues such as oracle model selection and lack of error bars reinforce conditional acceptance but are less decisive.","tokens_in":17026,"tokens_out":6445,"duration_ms":79438,"concrete_test":"Re-run the 1D benchmarks (1A,1B,2A,2B,3A,3B) and 2D Burgers with a self-contained pipeline: (1) obtain PD only from the AN indicator; (2) fit Φ(x,t)=0 through detected PD points by least squares; (3) set s = -Φ_t/|∇_x Φ|; (4) retrain with the published hyperparameters and report MSE All. If the CLINN-vs-PINN improvement drops substantially, or if the Buckley-Leverett double contact is not detected so LRH is inactive there, the published gains rely on analytic shock-speed information. Releasing the exact code path that constructs s would also resolve the ambiguity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is LCHAR = wIMLIM + wBDLBD + wRHLRH (Eq. 3.1). The LRH term, simplified in Eq. (3.2), penalizes |(f(û_L)-f(û_R))/(û_L-û_R) - s(xj,tj)| at detected discontinuity points. Evaluating this requires s(xj,tj), the shock speed. The paper says only that 'Φ(x,t) can be obtained through interpolation' (Sec. 3.1) and never specifies how s is computed from the trained network or from the AN-detected PD. In every numerical experiment, the exact discontinuity surface is supplied explicitly: Eq. (4.6) for (1A), Eq. (4.26) for the 2D case, and exact Riemann solutions for other benchmarks. If the implementation takes s from these analytic Φ, then LRH is a supervised term using the exact shock speed/location, rather than a self-contained conservation-law constraint. The reported 87.4–99.2% improvement over PINN would then reflect privileged information unavailable in a genuine forward solve, directly undermining the 'practical PINN variant' claim. This also explains the acknowledged double-contact limitation: for Buckley-Leverett (3A), λ(0)=λ(1)=0, so λL−λR=0 and the AN indicator cannot identify the contact; unless exact Φ is used, LRH would be inactive exactly where the solution has its discontinuity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLINN, a PINN variant for scalar conservation laws that augments the usual governing-equation and initial/boundary losses with three conservation-law-informed terms: an implicit-solution loss LIM, a boundedness loss LBD, and a Rankine-Hugoniot jump-condition loss LRH. It also uses the Feng–Liu artificial-neuron indicator to detect discontinuities and a residual-based adaptive refinement (RAR) scheme to reweight collocation points near shocks. The method is tested on 1D inviscid Burgers, LWR traffic flow, Buckley-Leverett, and a 2D Burgers problem, reporting large MSE reductions relative to PINN (up to 99.2%) and generally better accuracy than IFNN and PINN-WE.","tokens_in":17402,"tokens_out":6361,"duration_ms":72990,"significance":"If the method were self-contained, the reported gains would be practically valuable for approximating discontinuous solutions of scalar conservation laws. The paper gives a detailed, reimplementable experimental setup, compares against three baselines, includes an RAR ablation, and honestly acknowledges the AN indicator's limitation for double contact discontinuities. However, the central novelty—the Rankine-Hugoniot loss—appears to rely on externally supplied exact shock-speed information in every benchmark, which would make the comparison with PINN unfair and undermine the claim of a practical forward solver. The evaluation also uses test-set oracle checkpoint selection with no error bars. These issues are load-bearing for the paper's central claim.","major_comments":[{"comment":"The Rankine-Hugoniot term LRH is evaluated at detected discontinuity points and penalizes |(f(û_L)-f(û_R))/(û_L-û_R) - s(xj,tj)|. The manuscript never specifies how s(xj,tj) is obtained from the trained network or from the AN indicator. In all numerical experiments the exact discontinuity surface is available and stated: Eq. (4.6) for (1A), Eq. (4.26) for the 2D case, and exact Riemann solutions for the other cases. If s is taken from these analytic expressions, LRH is a supervised term using analytic shock data. If, alternatively, s is computed from the same network values û_L and û_R, the term is identically zero and imposes no constraint. Thus the central loss term either requires privileged information or is vacuous. A practical forward PINN variant needs an autonomous way to obtain s; the current text does not provide one.","section":"§3.1, Eq. (3.1)–(3.2)"},{"comment":"The AN indicator is acknowledged to fail on double contact discontinuities because λL−λR vanishes identically. The Buckley-Leverett case (3A) has exactly such a double contact, since λ(1)=λ(0)=0 for the flux in Eq. (4.15). Yet CLINN is reported to outperform on (3A) (Table 4), and Fig. 9(h) is described as agreeing with exact shock positions. This is internally inconsistent unless the discontinuity set PD is taken from the exact solution rather than from the indicator. Please clarify how PD was constructed for this case and, more generally, what 'Φ(x,t) can be obtained through interpolation' means in §3.1.","section":"§3.2, test case (3A)"},{"comment":"The evaluation selects checkpoints using the exact solution: 'We save each model that achieves the smallest MSE between the predicted and the exact solutions after 1000 epochs.' Because the same test MSE is the evaluation metric, the reported MSEs in Tables 2–5 and the improvement ratios of Eq. (4.1) are optimistically biased. No multi-seed statistics or error bars are provided. The comparison would be much more convincing with fixed-epoch models, early stopping on a validation set, or repeated independent runs.","section":"§4.1, 'Comparisons'"},{"comment":"The phrase 'enforcing exact conservation properties' is an overclaim. The boundedness term LBD and the Rankine-Hugoniot term LRH are soft penalties in the loss; they do not enforce exact conservation either locally or globally. The paper provides no conservation-error measurement, such as the evolution of the spatial integral of the solution. Please revise the wording and, ideally, report a conservation-error metric to support the claim.","section":"Abstract and §5"}],"minor_comments":[{"comment":"There is a typo in the loss definition: 'wICLBC' should presumably be 'wBCLBC'. Please correct.","section":"Eq. (2.12)"},{"comment":"Minor language issues: 'numeral oscillations' should be 'numerical oscillations'; 'Two-Dimenstional' should be 'Two-Dimensional'; 'Comparation' in table captions should be 'Comparison'.","section":"Abstract / §4.3 / tables"},{"comment":"The notation HardTanh²(û; inf u0, sup u0) is unclear. If it means the squared distance to the interval [inf u0, sup u0], please define it explicitly, because the current notation could be read as the square of the HardTanh output.","section":"§3.1, LBD definition"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is whether LRH is supervised by exact shock data. I would ask the authors to provide the actual construction of PD and s, and to run an experiment where s is estimated from the network's detected shock path (e.g., finite differences of the detected discontinuity location). If this cannot be done, the central practical claim should be withdrawn or substantially narrowed. The oracle checkpoint selection should also be fixed in any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read of CLINN. The combination is genuinely new: implicit-solution loss plus boundedness plus Rankine-Hugoniot jump term, with an AN-based shock indicator driving an RAR reweighting scheme. That packaging, tested across Burgers, LWR, and Buckley-Leverett with convex, concave, and non-convex fluxes, is a solid piece of engineering work. The authors are also honest where it counts: they admit the AN indicator misses double contact discontinuities, and they report that CLINN is slightly worse than IFNN on case (1A). The method description is detailed enough to reimplement, and the ablation \"CLINN w/o RAR\" helps isolate the adaptive weighting contribution.\n\nBut there is a load-bearing problem. The Rankine-Hugoniot term needs the shock speed s(x_j,t_j). The paper only says that the discontinuity surface Φ \"can be obtained through interpolation\" and never specifies how s is computed from the network or from the detected points. In every numerical experiment, the exact Φ is supplied explicitly (Eq. 4.6, 4.26, etc.). If the implementation uses that exact Φ to evaluate s, then LRH is a supervised term using the exact shock location and speed. The reported 87%–99% improvement over PINN would then be the result of privileged information, not a self-contained conservation-law constraint. The acknowledged double-contact failure is consistent with this: for case (3A), the AN indicator provably cannot identify the contact, so the only way LRH could be active is through the exact Φ. This doesn't make the paper useless, but it fundamentally changes what the headline numbers mean.\n\nOther problems are more routine. There are no repeated runs or error bars; the best checkpoint is selected by test MSE against the exact solution; and the phrase \"exact conservation\" overstates what a soft loss penalty provides. The comparison omits recent weak-form and thermodynamically consistent PINN baselines, though the three baselines they do use are appropriate.\n\nMy overall verdict: this is a serious paper with a real idea but a central methodological gap. The right fix is either to rerun without exact shock data (estimating s from the network or from a separate detection step) or to reframe CLINN as using exact shock information as an additional supervised input. As written, the forward-solver claim is not supported.\n\nI'd send it to peer review rather than desk-reject: the combination of ideas merits careful referee scrutiny, and the flaw is fixable. I'd also want code and error bars before publishing. If I were working in this area, I'd cite it as an example of how easy it is to leak exact solution information into a supposedly physics-informed loss.\n\nBest,\n[Your name]","headline":"A promising loss-function cocktail for PINNs on conservation laws, but the headline gains likely rely on exact shock-speed data baked into the Rankine-Hugoniot term.","tokens_in":165,"tokens_out":2245,"would_cite":false,"duration_ms":60192,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M99","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a PINN trained with additional implicit-solution, boundedness, and Rankine-Hugoniot losses, plus adaptive refinement driven by an artificial-neuron shock indicator, approximates discontinuous scalar conservation-law so","keywords":["scalar conservation laws","physics-informed neural networks","Rankine-Hugoniot condition","shock detection","adaptive refinement","Burgers equation","Buckley-Leverett equation","traffic flow model"],"falsifier":"Train CLINN on a Riemann problem whose exact solution is a double contact discontinuity, for example a Buckley-Leverett initial condition arranged so the characteristic speeds on both sides coincide; the paper itself states the AN indicator provably misses such waves, so if the predicted shock location remains accurate anyway, the LRH term is doing less work than claimed, and if it drifts, the central claim fails for this wave type. A second check is to replace the interpolated Phi with the exact discontinuity surface and compare MSE: if the result changes materially, the method's success depe","tokens_in":16937,"feed_emoji":"🌊","tokens_out":7578,"duration_ms":80198,"temperature":0.7,"pith_summary":"CLINN is a recipe for training neural networks to solve scalar conservation laws with moving shocks. It starts from the standard physics-informed loss and adds three terms that encode mathematical facts every entropy solution obeys: the implicit relation u = u0(x - f'(u)t), the bound between the initial data's infimum and supremum, and the Rankine-Hugoniot jump condition across discontinuities. A two-neuron artificial-neuron indicator locates shocks, and an improved residual-based adaptive refinement scheme concentrates training weight near them. On benchmarks from Burgers, traffic flow, and Buckley-Leverett, the paper reports lower mean-squared error than PINN and its immediate variants, with the largest improvement a 99.2% MSE reduction. The practical claim is that conservation-law structure, not just the PDE residual, is what lets PINNs handle discontinuities.","feed_headline":"Shock-solving neural nets cut PINN error by 99.2 percent","feed_subtitle":"Adding conservation-law constraints pinpoints moving discontinuities in Burgers, traffic, and Buckley-Leverett benchmarks.","key_machinery":"The load-bearing object is a composite loss L = LPINN + wIM LIM + wBD LBD + wRH LRH. LIM is the implicit-solution residual (predicted u minus u0(x - f'(u)t)); LBD is the boundedness penalty implemented with HardTanh against the initial data's range; LRH is the Rankine-Hugoniot residual at detected discontinuity points, comparing the network's flux jump to the shock speed. The discontinuity set comes from the two-neuron AN indicator, and the improved RAR reweights points according to the governing-equation and implicit-solution residuals outside that set. The combination is what carries the argument: each term injects a different piece of conservation-law structure that the raw PDE residual c","core_discovery":"The paper claims that a loss function augmented with three conservation-law-specific terms—the implicit solution form, the boundedness constraint, and the Rankine-Hugoniot jump condition—turns a PINN into a reliable approximator of discontinuous entropy solutions. The implicit solution term penalizes deviations from u = u0(x - f'(u)t) where the solution is smooth; the boundedness term uses HardTanh to discourage values outside the initial range; and the jump-condition term enforces the shock speed computed from the flux jump at points flagged as discontinuities. On top of this, an artificial-neuron indicator marks shock cells and an improved RAR routine increases loss weights at high-residua","pith_inferences":["If the AN indicator's blind spot for double contact discontinuities is patched or replaced by a detector that includes characteristic-speed coincidence, the same loss formula should cover all wave types, since the Rankine-Hugoniot term itself does not require convexity.","Because the paper shows CLINN trails IFNN on the simple periodic Burgers case, an adaptive switch that activates LRH only when a shock is actually detected could avoid the optimization slowdown while preserving accuracy on harder cases.","The scalar-law construction suggests a route to systems: enforce the vector Rankine-Hugoniot condition componentwise along each wave family, though the paper leaves systems to future work.","A testable extension is to replace the unstated interpolation of the discontinuity surface Phi with an exact or independently computed surface and measure how much of the MSE gain comes from that construction rather than from the loss terms themselves."],"forward_implications":["For scalar conservation laws with convex, concave, and nonconvex fluxes, CLINN reproduces shock locations and solution profiles with lower MSE than PINN, IFNN, and PINN-WE across the reported benchmarks.","The boundedness loss suppresses spurious oscillations on the downstream side of shocks, visible in the LWR and nonconvex cases.","Combining the implicit-solution and Rankine-Hugoniot terms selects the physically admissible solution in regions where the implicit form is multivalued, such as the rarefaction-shock interaction in the LWR test.","The improved RAR reweights high-residual points near discontinuities, adding further error reduction after the first training phase; the ablation model without RAR performs worse after 5000 epochs.","The framework extends to 2D by applying the AN indicator via dimensional splitting, demonstrated on the 2D Burgers equation."],"supporting_citations":[{"why":"Supplies the base physics-informed loss that CLINN extends and the standard PINN baseline it is compared against.","marker":"[28]"},{"why":"Introduces the implicit solution form loss that CLINN adopts and extends.","marker":"[33]"},{"why":"Introduces the Rankine-Hugoniot jump-condition loss that CLINN extends.","marker":"[23]"},{"why":"Supplies the artificial-neuron shock indicator used to identify discontinuity points for LRH and RAR.","marker":"[12, 11]"},{"why":"Supplies the residual-based adaptive refinement idea for concentrating training near large residuals.","marker":"[25]"},{"why":"Supplies the gradient-annihilated weighting idea that the improved RAR adapts to emphasize under-resolved smooth regions.","marker":"[13]"},{"why":"Provides the conservation-law theory (implicit solution, boundedness, Rankine-Hugoniot condition, entropy condition) that motivates the added loss terms.","marker":"[4]"}],"fun_headline_variants":["Physics-aware loss tames shock waves in neural nets","99.2% MSE cut: CLINN pins down discontinuities","Exact conservation meets deep learning: shock solver wins","CLINN: 99.2% error drop on shock benchmarks","Conservation-law training sharpens neural nets at shocks"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the artificial-neuron indicator tags every real discontinuity, but it provably misses double contact discontinuities, and the discontinuity surface Phi used in the Rankine-Hugoniot loss has no explicit construction from the trained network, so if detection or interpolation is wrong the jump-condition loss is enforced at the wrong places and the reported gains can degrade.","fun_headline_variants_meta":{"raw":{"variants":["Physics-aware loss tames shock waves in neural nets","99.2% MSE cut: CLINN pins down discontinuities","Exact conservation meets deep learning: shock solver wins","CLINN: 99.2% error drop on shock benchmarks","Conservation-law training sharpens neural nets at shocks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3614,"prompt_tokens":713,"completion_tokens":2901,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":2827}},"tokens_in":457,"tokens_out":2901,"duration_ms":21980,"temperature":1.0,"reasoning_tokens":2827,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:51:49.658476+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CLINN on a Riemann problem whose exact solution is a double contact discontinuity, for example a Buckley-Leverett initial condition arranged so the characteristic speeds on both sides coincide; the paper itself states the AN indicator provably misses such waves, so if the predicted shock location remains accurate anyway, the LRH term is doing less work than claimed, and if it drifts, the central claim fails for this wave type. A second check is to replace the interpolated Phi with the exact discontinuity surface and compare MSE: if the result changes materially, the method's success depe","supporting_citations":[{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"Supplies the base physics-informed loss that CLINN extends and the standard PINN baseline it is compared against."},{"cited_title":"Zhang, T","cited_arxiv_id":null,"evidence_quote":"Introduces the implicit solution form loss that CLINN adopts and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Rankine-Hugoniot jump-condition loss that CLINN extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the residual-based adaptive refinement idea for concentrating training near large residuals."},{"cited_title":"Ferrer-Sanchez, J","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-annihilated weighting idea that the improved RAR adapts to emphasize under-resolved smooth regions."},{"cited_title":"Chang and L","cited_arxiv_id":null,"evidence_quote":"Provides the conservation-law theory (implicit solution, boundedness, Rankine-Hugoniot condition, entropy condition) that motivates the added loss terms."}],"review_version":1}