{"id":"cb69ae46-4fac-44ec-aaa7-837b458349fe","arxiv_id":"2511.09048","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A projection layer can enforce linear and quadratic integral conservation in PINNs, but the quadratic projection formula as printed omits the discretization factor and therefore does not satisfy its own constraint.","lead":"This paper adds a projection step to physics-informed neural networks that forces the network output to keep a chosen integral quantity exactly equal to a target value. The method cuts conservation errors by orders of magnitude, but one printed formula is missing a scale factor and must be corrected before the guarantee holds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quadratic projection formula in Eq. (15)/(41) misses the Δx factor: substituting it into the stated constraint gives Δx Σ y_i² = Δx c(t), not c(t). The central guarantee for the quadratic integral is not established as written.","rationale":"I read the paper's central claim as a method-level guarantee: the projection layer should map an arbitrary PINN output to one satisfying the stated conservation constraint. The quadratic derivation is the key place where this must hold, and Eq. (15)/(41) fails the substitution test. This is not a matter of interpretation or external consensus; it is an internal inconsistency. The reader's rationale already calls this out, although the 'weakest_assumption' field emphasizes the discrete-grid/c(t) approximation; I see the algebra error as more load-bearing because even on the discrete grid the printed formula does not conserve. The reported experiments (Errorc ~1e-6) suggest the implementation may silently use the corrected factor, but the manuscript does not provide it, so the central claim is unsupported as written. I agree with the reader's REJECT conclusion; since my read does not change that verdict, I mark the adjustment as UNCHANGED. A clean fix would be to correct Eq. (15)/(41)/(42), rerun or confirm the quadratic experiments, and restate the guarantee as discrete-quadrature conservation.","tokens_in":12619,"tokens_out":6603,"duration_ms":62505,"concrete_test":"Analytically recompute the quadratic projection: with n=256, Δx=1/128, pick any u and evaluate Eq. (15) then Δx Σ y_i^2; it equals Δx·c(t) ≈ c(t)/128. The correct solution is y_i = u_i sqrt(c/(Δx Σu²)). Then rerun the 1D Advection 'Quadratic' experiment with the corrected formula, or inspect the released code to see whether the implementation includes the Δx factor. If the code matches the printed Eq. (15), the reported Errorc values near 1e-6 cannot be reproduced; if it uses the corrected factor, the equations in §3.7/§A.11 must be amended.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that PINN-Proj 'guarantees the conservation of the linear and quadratic integrals.' For the quadratic integral, §3.7/§A.11 pose min_y ||u-y||^2 s.t. Δx y^T y = c(t) (Eq. 14) and derive y_i = u_i sqrt(c(t)/Σ_i u_i^2) (Eq. 15, repeated as Eq. 41). Substituting the printed y into the constraint yields Δx Σ_i y_i^2 = Δx·c(t), because Σ_i y_i^2 = c(t). The printed projection therefore does not satisfy the constraint it is meant to enforce; with the grid used here (Δx=1/128) the conserved quantity misses by a factor of 128. The KKT system actually implies y_i = u_i sqrt(c(t)/(Δx Σ_i u_i^2)); the missing Δx propagates to the continuous extension in Eq. 42. The linear projection (Eq. 12) and the combined projection (Eq. 17) are internally consistent, so the flaw is specific to the quadratic-alone result, which is one of the three headline configurations in Table 3. This is an internal algebraic inconsistency in the central derivation, not a disagreement with prior work. Secondary limitations about discrete quadrature and FDM-derived c(t) would remain even after fixing this, but the missing Δx already blocks the guarantee as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PINN-Proj, a hard-constraint layer for physics-informed neural networks that projects the network output onto a manifold where selected integral quantities (linear integral, quadratic integral, or both) are conserved. The projection is derived by solving constrained least-squares problems with Lagrange multipliers. The method is tested on 1D/2D advection, wave, Korteweg-de Vries, and reaction-diffusion equations, comparing solution error and conservation error against a vanilla PINN and a soft-constraint variant (PINN-SC). The authors report that PINN-Proj reduces conservation error by three to four orders of magnitude and marginally improves solution error, and they analyze Hessian spectral densities to argue that the projection improves the conditioning of the loss landscape.","tokens_in":13012,"tokens_out":4686,"duration_ms":44666,"significance":"If the claims hold, the paper offers a practically useful way to enforce hard conservation constraints in PINNs using differentiable closed-form projections, which is an active area of research. The method is simple, does not require iterative projection steps, and is demonstrated on several PDEs. The paper also provides a useful Hessian-spectrum analysis that connects the projection to improved conditioning. However, the central algebraic derivation for the quadratic projection contains an error that undermines the main guarantee, and the 'guarantee' is stated only for a discrete quadrature, not for the continuous integral.","major_comments":[{"comment":"The quadratic projection formula is algebraically incorrect. Solving the KKT system in Eq. (36)-(39) gives y_i = u_i/(1+λΔx) and, after substituting into Δx yᵀy = c(t), one obtains 1/(1+λΔx) = sqrt( c(t) / (Δx Σ_i u_i²) ). The paper instead writes 1/(1+λΔx) = sqrt( c(t) / Σ_i u_i² ), omitting the factor Δx in the denominator. Consequently, Eq. (15) and Eq. (41) do not satisfy the stated constraint: substituting the printed formula yields Δx Σ_i y_i² = Δx c(t), not c(t). With the experimental Δx = 1/128, this is a factor-of-128 discrepancy. This is a load-bearing error for the claim of guaranteed quadratic conservation. Please correct the formula, update the continuous extension in Eq. (42), and verify whether the implementation and the reported Errorc_Q values in Table 3 correspond to the corrected or the printed formula.","section":"Sec. 3.7, Eq. (15); Sec. A.11, Eqs. (40)-(42)"},{"comment":"The abstract and introduction state that the method 'guarantees the conservation of the linear and quadratic integrals.' As written, however, the projection enforces the discrete quadrature conditions Δx Σ_i y_i = c(t) and Δx Σ_i y_i² = c(t), not the continuous integrals ∫u dx and ∫u²dx. The discrete-to-continuous gap is a discretization error that is acknowledged only indirectly in the discussion. The guarantee should be phrased as a discrete conservation guarantee, with the continuous guarantee holding only in the limit Δx→0. This is a conceptual point that affects the interpretation of all experimental claims.","section":"Secs. 3.4-3.8, 5; Abstract"},{"comment":"For non-conserved systems (e.g., reaction-diffusion), the target c(t) at collocation times is estimated by linear interpolation from FDM-derived values using Eq. (22). The projection then enforces this estimated value, so the reported conservation error for non-conserved systems includes the interpolation error, which the paper itself notes in Sec. 5. This means the 'guarantee' is conditional on the quality of the c(t) approximation. This limitation should appear in the main text, not only in the discussion, and the effect of the interpolation error should be quantified or at least bounded.","section":"Sec. 3.9, Eq. (22); Sec. 5"}],"minor_comments":[{"comment":"The notation 1_n^T uθ(x,t) Δx is confusing: it suggests a scalar Δx multiplying a scalar dot product. Please define uθ(·) as a vector over spatial indices and write the sum explicitly or use a clearer vector notation.","section":"Sec. 3.6, Eq. (10)"},{"comment":"The combined projection formula in Eq. (17) is visually unreadable due to the nested summations and missing parentheses. Adding brackets around the summation terms would improve clarity.","section":"Sec. 3.8, Eq. (17)"},{"comment":"The bolding in Tables 2 and 3 is described as 'lowest error for that configuration,' but it is not clear whether bold comparisons are across models within the same PDE and conserved quantity, or across conserved quantities within a model. Please specify.","section":"Sec. 4, Tables 2-3"},{"comment":"The Hessian spectral density plots are informative, but the axes are not labeled and it is unclear whether the x-axis is linear or logarithmic. Please add axis labels and specify the kernel bandwidth used for smoothing.","section":"Sec. 5, Figure 2"},{"comment":"The claim that the projection 'improved convergence' is based on Hessian eigenvalue clustering. This is suggestive but not a direct proof of faster convergence. Consider supplementing with convergence curves or training loss trajectories.","section":"Sec. 5, 'PINN Training and Convergence'"}],"recommendation":"major_revision","confidential_remarks":"The missing Δx in the quadratic projection is a serious algebraic slip in a central part of the paper. If the code used the printed formula, the reported conservation-error results are not meaningful; if the code used the correct formula, the text and derivations must be corrected and the discrepancy acknowledged. Either way, the authors need to verify the implementation against the corrected formula. The conceptual issue of discrete-versus-continuous conservation also needs tightening. These are fixable within the scope of a major revision, and the overall idea remains promising if corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the stress-test note is correct. Eq. (15) and Eq. (41) give y_i = u_i sqrt(c(t)/Σ u_i²), but plugging that into the stated constraint Δx y^T y = c(t) yields Δx c(t), not c(t). The correct scaling is sqrt(c(t)/(Δx Σ u_i²)). With Δx=1/128, the printed formula misses by a factor of 128. Since that equation is the basis for the quadratic-conservation guarantee, the central claim is not supported as written. This is not a matter of interpretation.\n\nWhat’s actually new: the combined linear-plus-quadratic projection, Eq. (17)/(58), appears to be a genuine closed form not in the cited prior work. The linear projection is standard hyperplane projection, and the quadratic is norm scaling, but the joint version is a nice contribution. The paper is clearly written, the experimental setup is thorough, and the Hessian conditioning analysis is a good addition. The authors also honestly discuss discretization and solver-error sources.\n\nSoft spots, in proportion: the missing Δx is the big one, and it is load-bearing, not a typo in a footnote. It needs to be fixed and the corrected formula verified. Second, the “guarantee” is only for the discrete quadrature sum, not the continuous integral; the paper acknowledges this but the abstract could easily oversell it. Third, the Errorc improvements are by construction — the projection enforces the discrete conservation exactly, so comparing Errorc between PINN-Proj and the others is not an independent test. The real evidence for value is the Erroru comparison (marginal gains) and the conditioning shift, which are modest. Fourth, for non-conserved systems, c(t) is linearly interpolated from FDM data; if that data is wrong, the projection enforces the wrong law. These are limitations, not fatal flaws, but they should be stated clearly.\n\nWho this is for: people doing hard constraints in PINNs and scientific-ML surrogate modeling. A reader who needs mass or energy conservation as a hard layer might find this useful after the correction. I’d send it to peer review; the fix is small and the rest of the paper is solid. A referee should ask for the corrected derivation, a sensitivity check on the interpolation, and a more prominent statement that the guarantee is discrete. If those are addressed, I’d be okay with it.","headline":"The quadratic projection formula is off by a factor Δx, invalidating the headline guarantee as written, but the fix is small and the rest of the paper is solid enough to send to review.","tokens_in":13443,"tokens_out":3704,"would_cite":false,"duration_ms":34365,"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":"A projection layer that pins a PINN's output to the conservation manifold reduces integral-conservation error by three to four orders of magnitude while slightly improving solution accuracy.","keywords":["physics-informed neural networks","conservation laws","projection layer","hard constraints","quadrature","constrained optimization","loss landscape conditioning","PDE solving"],"falsifier":"Take a PDE with a known analytic conserved quantity, train PINN-Proj on a coarse collocation grid, then compute the continuous integral (e.g., ∫u² dx) with a much finer quadrature; if the continuous error does not drop alongside the discrete conservation error, the 'guarantee' is an artifact of the grid. Alternatively, feed the projection a deliberately wrong c(t) trajectory and observe whether the solution error increases; a correct mechanism should respond to a wrong constraint.","tokens_in":12523,"feed_emoji":"🎯","tokens_out":8901,"duration_ms":77460,"temperature":0.7,"pith_summary":"This paper shows that conservation of linear and quadratic integral quantities can be enforced exactly in a physics-informed neural network (PINN) by projecting the raw network output onto the constraint set at every evaluation. The projection is the closed-form solution of minimizing the squared distance to the conserved manifold, derived with Lagrange multipliers, and it is differentiable, so it can be trained end-to-end. Across five PDEs in one and two dimensions, the projected network reduces conservation error by three to four orders of magnitude compared to a soft penalty, while the PDE solution error is at worst unchanged and at best reduced. The authors also report that the projection improves the conditioning of the loss landscape, making gradient descent converge in fewer epochs. The method is proposed as a general recipe: any integral quantity with a tractable differentiable projection can be conserved this way.","feed_headline":"Projection layer guarantees conservation on the grid in PINNs","feed_subtitle":"Forcing conservation as a hard constraint cuts error by three to four orders over soft penalties.","key_machinery":"The central object is the projection layer. For the linear integral, the projection subtracts the current spatial mean from the output and adds the target mean c(t)/(nΔx). For the quadratic integral, it rescales the entire output vector by the square root of the ratio of the target quadratic sum to the current one. For the joint constraint, it first centers the output by subtracting its mean, rescales the centered vector to satisfy the quadratic target, and adds back the mean term set by the linear target. Each formula is the closed-form solution of a constrained least-squares problem obtained with Lagrange multipliers. The layer operates on the discretized output on a uniform quadrature gri","core_discovery":"The paper claims that a projection layer can guarantee conservation of the linear integral, the quadratic integral, and both jointly, in the sense that the only remaining errors are those from the data and from machine precision. The projection is computed by solving a constrained non-linear optimization problem — minimize the L2 distance between the network output and the set of functions with the prescribed integral — and the solution is a simple closed form: a constant shift for the linear constraint, a uniform scaling for the quadratic constraint, and a shift-around-the-mean followed by a scaling for the combined constraint. Because the projection is applied to the discretized output on","pith_inferences":["The paper leaves implicit that the same projection recipe could be applied to other integral functionals whenever the constrained least-squares problem has a closed form; this would turn the method into a general library for hard conservation in neural PDE solvers.","The reported Hessian improvement suggests projection acts as a conditioning regularizer; a controlled study on stiffer PDEs could test whether the benefit grows precisely where soft constraints struggle most.","Because the projection is a differentiable map applied at inference time, it could serve as a post-processing step to restore conservation in an already-trained PINN without retraining.","A stress test on a coarse grid with a high-resolution continuous integral would reveal whether the discrete guarantee translates to the continuous conservation law, which is the quantity physics actually cares about."],"forward_implications":["On all five PDEs tested, PINN-Proj reduces conservation error for the linear and quadratic integrals by three to four orders of magnitude compared with the soft-constrained baseline.","PINN-Proj achieves the lowest PDE solution error on 11 of 15 configurations, with margins under one order of magnitude.","Projected models converge in the fewest epochs across all PDEs, and the Hessian spectrum is more tightly clustered, implying a better-conditioned loss landscape.","The method extends to 2D domains and to non-conserved systems by interpolating the time-dependent target c(t), so it applies beyond exactly conserved systems.","Because the projection is differentiable and applied at inference time, it can be inserted as a layer in any PINN-like architecture without changing the training objective."],"fun_headline_variants":["Projection guarantees integral conservation in PINNs","Hard constraint projection cuts PINN conservation error by 3-4 orders","PINN-Proj: exact linear and quadratic integral conservation","Projection layer locks in integrals in PINNs","Projection enforces physical conservation laws in PINNs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The guarantee is defined on the discrete quadrature grid — it enforces the summed quantity, not the continuous integral — and for non-conserved systems the target c(t) is itself interpolated from data, so if that target is wrong the projection enforces the wrong conservation law.","fun_headline_variants_meta":{"raw":{"variants":["Projection guarantees integral conservation in PINNs","Hard constraint projection cuts PINN conservation error by 3-4 orders","PINN-Proj: exact linear and quadratic integral conservation","Projection layer locks in integrals in PINNs","Projection enforces physical conservation laws in PINNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00049,"raw_usage":{"total_tokens":2212,"prompt_tokens":673,"completion_tokens":1539,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":417,"completion_tokens_details":{"reasoning_tokens":1461}},"tokens_in":417,"tokens_out":1539,"duration_ms":12285,"temperature":1.0,"reasoning_tokens":1461,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:41:29.157978+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a PDE with a known analytic conserved quantity, train PINN-Proj on a coarse collocation grid, then compute the continuous integral (e.g., ∫u² dx) with a much finer quadrature; if the continuous error does not drop alongside the discrete conservation error, the 'guarantee' is an artifact of the grid. Alternatively, feed the projection a deliberately wrong c(t) trajectory and observe whether the solution error increases; a correct mechanism should respond to a wrong constraint.","supporting_citations":[],"review_version":1}