{"id":"01bf02fe-b523-44ec-9e15-8dc3eec3eb54","arxiv_id":"2509.10945","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A known PINN variant with boundary-layer subnetworks is applied to more singular perturbation problems, but accuracy is only reported via training loss, not solution error.","lead":"This paper applies a 2024 method called Chien-Physics-Informed Neural Networks (C-PINNs) to a wider set of singularly perturbed equations, including coupled and two-dimensional cases. The authors report lower training losses than standard PINNs, but they do not measure actual solution errors and provide no code or data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy claim rests on training-loss values, but for the reported ε and 600 collocation points the loss cannot constrain the exponential boundary-layer terms, and no true solution error is reported.","rationale":"The reader's weakest_assumption is that training loss is equated with solution accuracy. I agree, and the more specific mechanism is severe: even as a training objective, the loss cannot constrain the boundary-layer part of the C-PINN ansatz for the reported ε and 600 collocation points. Because the exponential layer terms underflow at all interior collocation points, the inner-network parameters have essentially no gradient from the PDE residual; only the boundary-condition loss at x=0 or x=1 touches them. Thus a low reported loss (e.g., 7.7e-3 in §3.2.1) can coexist with an arbitrarily wrong layer profile. Since the exact solutions are known for every example and the paper still omits error tables, the central claim is unsupported. The post-hoc addition of an auxiliary inner network at y=1 in §3.3 further indicates the architecture was adjusted after observing failure. There is no code, data, or formal verification to independently corroborate the numbers. The proposed test—retraining one coupled example and computing true errors on a layer-resolving grid—would settle whether the loss comparison translates into actual accuracy. If the true errors are large or C-PINN is not clearly better than PINN, the paper should be rejected; if errors are small, a revision with error metrics would be needed. I therefore see no reason to alter the reader's reject verdict.","tokens_in":12121,"tokens_out":9632,"duration_ms":103574,"concrete_test":"Retrain the §3.2.1 coupled convection-diffusion C-PINN exactly as specified (ε=1e-7, μ=1e-5, 600 uniform collocation points, 7000 Adam epochs, lr=5e-4) and, under identical settings, a standard PINN. Then compute the true error of both models against the known exact solution on a grid of 10^5 points clustered near x=0 (e.g., x=(k/N)^2 for k=0..N plus uniform points on [0.01,1]), reporting max-norm and L2 error for u1 and u2. If C-PINN's max error is not at least an order of magnitude below PINN's, or if its max error exceeds 1e-2 while the reported loss is ~7.7e-3, the training loss is not a valid accuracy measure and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims C-PINNs 'provide a more accurate solution' than conventional methods. The only quantitative evidence is training-loss curves (Tables 1, 3, 4, 6; Tables 2 and 5 have no baseline). Section 3 states 'The model's accuracy is evaluated using the L2 loss, calculated as the MSE'—i.e., residual loss (6), not solution error. This proxy is not merely weak; for the stated parameters it cannot see the boundary layers. With 600 uniformly distributed collocation points on [0,1], the nearest interior point to a boundary is at distance ≈1.7e-3. The reported layer widths are O(ε)=1e-7 (convection-diffusion) and 1e-5/1e-10 (reaction-diffusion), so at every interior collocation point the exponentials exp(-x/ε) or exp(-(1-x)/ε) are underflow/zero (and are clamped to exp(-20) by the safe-exp defined in §3.2). The inner-network terms in (7), (8), and the coupled-system ansätze therefore contribute only at the boundary points; the residual loss is minimized by the outer network alone and provides no gradient signal that trains the layer subnetworks. The paper never reports true L2/L∞ errors against the known exact solutions; the only error mentions are a vague 'global relative error' in §3.3 and unlabeled absolute-error plots. Hence the comparative accuracy claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript applies Chien-PINNs (C-PINNs), a composite neural-network ansatz combining an outer network with exponentially weighted inner subnetworks, to singularly perturbed convection-diffusion, reaction-diffusion, and coupled systems in one and two dimensions. For each test problem, the source term is chosen from a known analytical solution, and the paper reports training-loss curves comparing C-PINN with standard PINN (or standalone C-PINN loss). The central claim is that C-PINNs 'provide a more accurate solution' than conventional methods. The paper includes algorithms and solution plots, but no true solution-error metrics are reported.","tokens_in":12508,"tokens_out":10423,"duration_ms":119379,"significance":"The composite-asymptotic ansatz is a plausible idea for problems with known boundary-layer structure, and the paper extends it to coupled and 2D systems, which is a useful test bed. Credit is due for giving explicit ansatz formulas and algorithmic pseudocode. However, the central accuracy claim is not established: all quantitative evidence is residual/boundary loss, and for the reported parameter ranges the layer terms are numerically invisible at interior collocation points. If corrected with layer-adapted collocation and true error metrics, the results could be a useful benchmark; in the current form, the claims outrun the evidence.","major_comments":[{"comment":"The paper measures accuracy by the residual/boundary loss (6), not by comparing the computed solution with the known exact solution. Section 3 states \"The model's accuracy is evaluated using the L2 loss, calculated as the MSE\", conflating the training objective with solution error. For singular perturbation problems, small residual loss does not imply small solution error, especially when layer terms are invisible to the residual. Tables 1, 3, and 4 therefore do not support the claim of \"more accurate solution\"; Tables 2 and 5 have no baseline. Please report L2/L∞ errors against the exact solutions.","section":"Section 3, Tables 1-6, Eq. (6)"},{"comment":"With 600 uniformly distributed collocation points on [0,1], the nearest interior point is roughly 8e-4 from a boundary. For ε=1e-5 to 1e-10, exp(-x/ε) is below machine precision at all interior points, and the safe_exp clamp (-20) makes these terms constants (~2e-9). Thus the inner subnetworks contribute almost nothing to the residual loss and receive no gradient; the reported loss minimization is effectively performed by the outer network alone. No conclusion about the composite C-PINN structure can be drawn from these runs. Use layer-adapted sampling or larger ε, and verify that the inner-network parameters actually change during training.","section":"§3.1-§3.3, Eqs. (7)-(8), §3.2 safe_exp"},{"comment":"The \"auxiliary inner network around y=1\" is an ad hoc modification added after the model \"struggled\" near y=1, with no theoretical justification and only a vague claim that \"global relative error\" was reduced. More generally, §2.4 requires knowing boundary-layer locations and thicknesses in advance, and the exponentials in Eqs. (7)-(8) are chosen to match the exact solutions. The experiments therefore demonstrate fitting known layer shapes rather than discovering layers. This limitation should be stated prominently, and at least one example with unknown or not-pre-specified layer structure should be included if generality is claimed.","section":"§3.3 Example 1, §2.4"},{"comment":"The coupled-system ansatz is written as a single scalar u(x) = outer_nn1 + inner_nn10·safe_exp(-x/ε) + outer_nn2 + inner_nn20·safe_exp(-x/µ), which is ambiguous for a two-component solution u=(u1,u2). Algorithm 2 instead defines separate uθ1 and uθ2, each with left and right layer terms, whereas the analytical solution has layers only at x=0. The actual network output structure is therefore not reproducible as written. Please specify exactly which ansatz is used for u1 and u2.","section":"§3.2.1, Eqs. (12)-(13)"}],"minor_comments":[{"comment":"p1, p2, and δ1 are used before being defined; notation for layer thickness varies among δ1, δ(ε), and ∂(ε) in Section 2.4.","section":"Section 2.2, Eq. (4)"},{"comment":"The soft-residual weight w(x)=e^{-λ|R(x)|} is described, but λ=0.8 is mentioned only once and no experiment isolating its effect is shown; the loss-weight parameters λ_Dε, λ_B, λ_I in Eq. (6) are never specified.","section":"§3.1.1"},{"comment":"At epoch 4500 the C-PINN loss jumps to 0.001955 from ~2.6e-5 at neighboring epochs, contradicting the text that C-PINN \"consistently achieves a lower loss across all epochs\".","section":"Table 1"},{"comment":"The term \"L2 loss\" is used for the residual MSE; this should be relabeled to avoid confusion with the L2 solution error.","section":"Section 3"},{"comment":"The name \"PI-PINN\" is used for the standard PINN baseline; the acronym is nonstandard and should be defined or replaced. Code and full hyperparameters are not provided, hampering reproducibility.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"I concur with the stress-test analysis: the reported loss values cannot validate the method for the quoted ε values, and the lack of true solution-error metrics is a serious gap. I am recommending major revision rather than reject because these deficiencies are addressable by re-running the experiments with layer-adapted collocation, reporting L2/L∞ errors against the known exact solutions, and clarifying the coupled-system ansatz. The authors should also address the circularity issue: the method assumes known layer structure, and the ad hoc inner network in 2D Example 1 further weakens the claim that C-PINN is a principled general method. If the revised version does not provide error tables and a non-hand-tuned example, rejection would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is a straightforward extension of Wang et al.'s C-PINN to coupled 1D and 2D singular perturbation problems. That extension is legitimate and the test problems (convection-diffusion, reaction-diffusion, coupled systems, 2D) are well chosen. The authors are also candid that C-PINN requires knowing boundary-layer locations and thicknesses in advance. But the central accuracy claim—that C-PINNs provide more accurate solutions than conventional methods—is not established.\n\nThe papers' quantitative evidence is training loss only. Tables 1, 3, 4, and 6 compare residual MSE between C-PINN and PINN, and Section 3 explicitly equates accuracy with the L2 loss computed as MSE. For singularly perturbed problems, a low residual loss does not imply small solution error. The paper never reports actual L2 or L-infinity errors against the known exact solutions, even though those exact solutions are available for every test case. That is a straightforward omission.\n\nThe stress-test note is correct, and it is worse than a generic 'loss is not error' point. With 600 uniform collocation points on [0,1] and epsilons from 1e-5 down to 1e-10, the exponential factors in the C-PINN ansatz are effectively zero at every interior collocation point (safe_exp clamps at exp(-20)). So the inner subnetworks are invisible to the residual loss; the loss is minimized by the outer network alone. The reported loss comparisons do not even show that the boundary-layer subnetworks are being trained, let alone that C-PINN is more accurate.\n\nThere is also a post-hoc element in 2D Example 1, where an auxiliary inner network at y=1 is added after the model struggled. The authors disclose this, but it means the experiment is not a clean test of the method. No code or data is provided, so the experiments cannot be reproduced.\n\nOn the positive side, the paper follows the prior C-PINN work closely and cites it properly. The idea of applying the method to coupled systems and 2D problems is worth a technical note, but not in this form. A revision that reports true solution errors, includes error bars, uses epsilons where the layer terms are actually visible to the collocation points, and releases code could be useful.\n\nWho is this for? Someone working on PINNs for singular perturbation problems might skim it as a catalog of test cases, but no one should rely on the accuracy claims. I would desk-reject this version.","headline":"Applies an existing C-PINN to coupled and 2D singular perturbation problems but never reports true solution errors, so the accuracy claim is unsupported—desk reject.","tokens_in":12985,"tokens_out":5173,"would_cite":false,"duration_ms":61980,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65L10","65L11","65L20","65L60","65L70","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims C-PINNs, which split a neural-network solution into a smooth outer part and boundary-layer inner parts joined by exponential weights, solve singularly perturbed convection-diffusion and reaction-diffusion problems more acc","keywords":["physics-informed neural networks","singular perturbation problems","boundary layers","composite asymptotic expansion","convection-diffusion equations","reaction-diffusion equations","coupled systems","C-PINN"],"falsifier":"For the reaction-diffusion example with ε = 1e-5, evaluate both the standard PINN and C-PINN solutions at 2000 uniformly spaced points in [0,1] after the same 9500 training epochs and compute the L2 and L∞ errors against the known analytical solution. If C-PINN's true error is not smaller than the standard PINN's—or if its residual loss is low while its pointwise error exceeds the PINN's—the paper's central accuracy claim collapses.","tokens_in":11960,"feed_emoji":"🧮","tokens_out":5021,"duration_ms":54749,"temperature":0.7,"pith_summary":"The paper tries to establish that a modified physics-informed neural network, called C-PINN, is more accurate than standard PINNs for singularly perturbed problems. Singular perturbation problems are hard because a tiny parameter multiplies the highest derivative, creating sharp boundary layers that standard networks smear out. C-PINN addresses this by approximating the solution as a smooth outer network plus inner networks that decay exponentially away from the boundary layers. The paper applies this construction to 1D and 2D convection-diffusion, reaction-diffusion, and coupled systems, and reports training losses orders of magnitude lower than standard PINNs. A sympathetic reader would care because this is a mesh-free path to resolving boundary layers without extreme grid refinement.","feed_headline":"Composite PINNs capture boundary layers standard PINNs miss","feed_subtitle":"By adding exponential-decay inner networks, training loss on singularly perturbed problems drops from 1e-2 to 1e-5 in tests.","key_machinery":"The central object is the composite neural network u(x) = u_outer(x;θ1) + u_inner0(x;θ2) exp(-p1(x)/δ1) + u_inner1(x;θ3) exp(-p2(x)/δ1), where p1 and p2 are signed distances to the suspected boundary-layer edges and δ1 is the layer-thickness scale tied to the perturbation parameter. The exponential weights force each inner network to contribute only inside its boundary layer, while the outer network handles the smooth remainder; all subnetworks are trained jointly by minimizing the mean squared error of the PDE residual and boundary conditions.","core_discovery":"On the paper's own terms, the discovery is that the composite asymptotic-expansion structure—outer neural network plus layer-specific inner neural networks weighted by exp(-distance/layer-thickness)—allows C-PINN to match the known analytical solutions of several singularly perturbed test problems with far lower residual and boundary-condition MSE than a single standard PINN. The reported losses drop from the order of 1e-2 to order 1e-5 or lower within roughly 10,000 epochs, including tiny perturbation parameters down to 1e-10. The authors claim this makes C-PINN better suited to singularly perturbed problems.","pith_inferences":["Editorial inference: the reported comparisons use training loss, not true solution error on a hold-out grid; a direct pointwise error test against the given analytical solutions would strengthen or qualify the accuracy claim.","Editorial inference: in the 2D example an auxiliary inner network was added near y=1 where no layer theoretically exists; this suggests C-PINN also acts as a flexible residual-reduction tool, so its benefit may not be limited to known layer locations.","Editorial inference: the exponential-decay weighting implicitly encodes a length scale, so C-PINN may be interpretable as a physics-informed network with built-in coordinate stretching; testing it on problems with interior layers or turning points would reveal how sensitive the method is to layer-location input.","Editorial inference: no runtime or cost comparison is given, so the claim of computational efficiency is not established; residual-loss speedups do not guarantee wall-clock or memory advantages."],"forward_implications":["If C-PINN delivers the reported convergence, it gives a mesh-free way to solve convection-dominated and reaction-dominated problems with layers of thickness O(1e-5) or thinner, without mesh refinement.","The same outer-plus-inner construction extends to coupled systems with two different perturbation parameters, as demonstrated by the coupled convection-diffusion and reaction-diffusion examples.","The 2D experiments indicate the architecture transfers to higher dimensions with layer-localized inner networks placed along coordinate directions.","Because layer location and thickness are inputs to the construction, the method presupposes prior knowledge of where boundary layers occur; the paper lists automatic layer detection as future work.","A direct corollary is that residual loss alone, not network width or optimizer choice, is the reported basis for the accuracy claim in these experiments."],"fun_headline_variants":["C-PINNs beat PINNs on tough boundary-layer problems","New neural net design tames singular perturbation equations","Exponential-layer PINNs slash error for steep boundary layers","C-PINNs: sharper solutions for singularly perturbed PDEs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper equates training loss (mean squared error of the PDE residual and boundary conditions) with solution accuracy, so if low residual loss can coexist with large true solution error, the reported loss comparisons do not prove C-PINN is more accurate.","fun_headline_variants_meta":{"raw":{"variants":["C-PINNs beat PINNs on tough boundary-layer problems","New neural net design tames singular perturbation equations","Exponential-layer PINNs slash error for steep boundary layers","C-PINNs: sharper solutions for singularly perturbed PDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1154,"prompt_tokens":666,"completion_tokens":488,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":410,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":410,"tokens_out":488,"duration_ms":5461,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:19:42.085344+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the reaction-diffusion example with ε = 1e-5, evaluate both the standard PINN and C-PINN solutions at 2000 uniformly spaced points in [0,1] after the same 9500 training epochs and compute the L2 and L∞ errors against the known analytical solution. If C-PINN's true error is not smaller than the standard PINN's—or if its residual loss is low while its pointwise error exceeds the PINN's—the paper's central accuracy claim collapses.","supporting_citations":[],"review_version":1}