{"id":"a82bcb50-e49f-4228-ba8b-363c73726ef3","arxiv_id":"2411.09329","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Learned SUPG stabilization parameters and adaptive indicator functions improve FastVPINNs L2 accuracy on two convection-dominated benchmarks.","lead":"This paper adds a SUPG-style stabilization term and a learned indicator function to hp-variational physics-informed neural networks for convection-dominated problems. The proposed architectures yield lower reported errors on two benchmark problems than prior PINN approaches, though the evaluation relies on in-sample hyperparameter tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported L2 errors use a uniform 100x100 grid that cannot sample layers of width O(eps)~1e-8, so the claimed accuracy gains may not measure layer resolution.","rationale":"The strongest claim is that learning tau and the indicator function yields a 'tremendous increase of accuracy' for convection-dominated problems. The reader correctly identifies in-sample hyperparameter tuning and best-epoch/best-run reporting as weaknesses. My stress-test found a more elementary problem: the error metric defined in (14) is a uniform 100x100 grid RMS. With eps=1e-8, Pout's outflow layers and Ppara's parabolic layers have widths about 1e-8 and 1e-4, respectively, both below the 0.01 grid spacing, so the reported errors exclude the layers that define the problem. The exact L2 contribution of a wrong layer alone is about 1e-4 for Pout and potentially 1e-2 for Ppara, which brackets or exceeds the claimed errors. Therefore the quantitative comparisons to baselines and the attribution of the improvement to the proposed mechanisms are not established by the data as reported. This is fixable by recomputing errors on a layer-resolving quadrature, so I keep the reader's conditional stance but for a more foundational reason: acceptance should require the layer-resolving error evaluation to confirm the claimed gains.","tokens_in":15355,"tokens_out":14055,"duration_ms":142842,"concrete_test":"Recompute the L2 error for the best configurations in Tables 5-8 using a layer-resolving error quadrature: for example, a 100x100 uniform grid enriched with geometrically spaced points at distances 1e-10 to 1e-2 from the outflow boundaries x=1 and y=1 (and, for Ppara, from y=0 and y=1), or high-order Gauss quadrature on a boundary-layer mesh with cell width about 1e-10 near those boundaries. If the layer-resolving L2 error is more than a factor of 2 larger than the reported discrete RMS, or changes the ranking among plain loss, learned tau, and adaptive indicator, then the paper's central quantitative accuracy claim is not supported by the current evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Equation (14) evaluates L2_err on a 100x100 equidistant grid with spacing about 0.01. For Pout and Ppara, eps=1e-8, so the outflow boundary layers have width O(eps)=1e-8 and the parabolic characteristic layers have width O(sqrt(eps))=1e-4, both far smaller than the grid spacing. No evaluation point lies inside these layers. Thus the discrete RMS values in Tables 5-8 measure only interior accuracy; the layers, which are the defining difficulty of convection-dominated problems, contribute essentially nothing to the reported numbers. A completely wrong layer of width delta and O(1) amplitude contributes at least O(sqrt(delta)) to the true L2 error, i.e. about 1e-4 for Pout and about 1e-2 for Ppara, which is comparable to or much larger than the claimed errors (1.0e-4 to 6.0e-5). Because the paper motivates the entire study by the need to resolve layers and attributes the 'tremendous increase of accuracy' to the learned tau and indicator, the reported metric is not a valid measure of the central claim. The hand-tailored indicator functions (16)-(17), chosen with knowledge of where the layers lie, further conflate the source of the improvement, but the metric problem is the more fundamental issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two extensions to the FastVPINNs framework for steady-state convection-dominated convection-diffusion-reaction problems in two dimensions. First, it adds a SUPG-style stabilization term to the variational loss and introduces a neural-network head that predicts a spatially varying stabilization parameter tau(x,y), scaled by a tunable factor tau_g. Second, it proposes to learn the exponents in a fixed functional family of indicator functions used for hard-constrained Dirichlet boundary conditions. The authors test both ideas on two manufactured-solution problems with epsilon = 1e-8 (Pout with outflow layers and Ppara with exponential and parabolic layers) and report L2 errors that are lower than those obtained with the plain variational loss and lower than PINN/hp-VPINN results from Frerichs-Mihov et al. (2024a). The central claim is that these mechanisms produce a 'tremendous increase of accuracy' for layer-dominated solutions.","tokens_in":15676,"tokens_out":4376,"duration_ms":42092,"significance":"If the reported accuracy gains are genuine and reproducible, the two architectural ideas would be a useful contribution to the growing literature on physics-informed neural networks for singularly perturbed problems. The learned-tau idea, in particular, is a natural way to avoid hand-tuning a stabilization parameter, and the adaptive indicator function addresses an underappreciated source of error in hard-constrained PINNs. The paper is clearly written and the experiments are easy to follow. However, the current evaluation protocol does not establish the central claim: the error metric in Eq. (14) cannot see the layers that motivate the whole study, and the hyperparameters are selected in-sample on the exact solutions of the test problems. The significance of the contribution therefore depends on whether these methodological issues can be fixed in a revision.","major_comments":[{"comment":"The L2_err metric in Eq. (14) is evaluated on a uniform 100x100 grid with spacing about 0.01. For Pout the outflow layers have width O(epsilon)=1e-8, and for Ppara the parabolic layers have width O(sqrt(epsilon))=1e-4; both are far below the grid spacing, so no test point lies inside the layers. Consequently, the values reported in Tables 5-8 measure only interior accuracy, while the layers are the defining difficulty of convection-dominated problems and the stated motivation of the paper. A completely wrong layer of width delta and O(1) amplitude contributes at least O(sqrt(delta)) to the true L2 error, roughly 1e-4 for Pout and 1e-2 for Ppara, which is comparable to or much larger than the claimed errors (1.0e-4 to 6.0e-5). The claim of a 'tremendous increase of accuracy' for layer resolution is therefore not supported by the reported metric. Please compute the true L2 error with layer-resolving numerical quadrature (or on a graded mesh aligned with the layers) and additionally report errors restricted to the layer regions.","section":"Eq. (14), Sections 4.2 and 5"},{"comment":"The hyperparameters tau, lambda, and tau_g are selected by grid search using the known exact solution of the same two test problems, as described in Section 4.3 and in the text accompanying Tables 5-6. This is in-sample selection, so the reported improvements over the plain loss are partially fitted to the test cases. Moreover, Section 4.3 states that the smallest error obtained during training is reported ('after each epoch and not necessarily after the final epoch'), which adds an optimistic best-epoch bias. Please adopt a validation protocol that does not use the test solution for tuning (e.g., a fixed a priori choice or a separate validation split) and report the error at a fixed final epoch, together with means and standard deviations over multiple independent runs.","section":"Section 4.3 and Section 5.1, Tables 3-6"},{"comment":"The adaptive-indicator results report a single L2_err value per configuration with no number of runs, no standard deviation, and no statement about whether the value is the best run or the average over runs. Since Section 4.3 demonstrates that the minimum error varies noticeably across runs even with identical parameters, the single-point comparisons in Tables 7-8 are insufficient to establish that the learned alpha, beta, and gamma consistently improve accuracy. Please add multi-seed statistics and explicitly state the selection rule for the reported values.","section":"Section 5.2, Tables 7-8"},{"comment":"The comparison with the literature is limited to the PINN and hp-VPINN results from Frerichs-Mihov et al. (2024a). No comparison with a standard stabilized finite element method is provided, even though the MUAS method is used to generate the reference solution for Ppara and is mentioned as the currently most promising approach in the introduction. The two manufactured-solution problems share a similar structure (square domain, axis-aligned layers), so the generality of the conclusions is not yet demonstrated. Adding a stabilized FEM baseline and at least one problem with a different layer configuration would substantially strengthen the evidence for the paper's central claim.","section":"Section 4.3 and Section 5, Tables 3-6"}],"minor_comments":[{"comment":"The column header 'best L2_err' conflicts with the text stating that the average L2_err values from 10 runs are presented; please clarify which value is actually reported in each table.","section":"Section 5.1, Tables 5-6"},{"comment":"For Ppara, the MUAS method is used to obtain a high-accuracy reference solution, but the accuracy of that reference (e.g., the mesh size and the resulting error) is never quantified; please state the grid and the estimated error of the reference solution.","section":"Section 4.1"},{"comment":"The extension function j(x) in the hard-constrained ansatz (6) is not given explicitly for Pout and Ppara; providing the formulas used for j(x) is necessary for reproducibility.","section":"Section 2 and Section 4.3"},{"comment":"The text states that the plotted values are averages over 5 runs, but the figure does not show error bars or the spread of the data; adding them would make the sensitivity analysis more informative.","section":"Figure 6"},{"comment":"The paper says that tau is set to zero on the boundary, but the indicator w(x) = tanh(50x)tanh(50y)tanh(50(1-x))tanh(50(1-y)) is only approximately zero there; please state whether quadrature nodes exactly on the boundary are excluded from the SUPG loss or whether the approximation is accepted.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope for a numerical analysis journal and the proposed architectural modifications are reasonable and potentially useful. The main obstacle to acceptance is that the numerical evaluation does not currently support the accuracy claims: the L2 metric cannot resolve the layers, and the hyperparameter selection is in-sample. Both issues are fixable with a revised evaluation protocol, which is why I recommend major revision rather than rejection. I would also encourage the authors to release code for the exact configurations used, since the FastVPINNs framework is described elsewhere and the present paper's contributions are purely numerical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful engineering paper, not a breakthrough. The genuinely new pieces are the network that outputs a spatially varying tau for a SUPG-style term in FastVPINNs, and the learned exponential indicator parameters for hard Dirichlet constraints. Both are natural adaptations of known ideas (Yadav-Ganesan stabilized FEM; Lu et al. hard constraints), but I have not seen them done exactly this way, and the internal comparisons are consistent: learned tau beats constant tau, adaptive indicator beats fixed. The exposition is clear, and the authors are honest that constant-tau selection is impractical and that best-run results are reported.\n\nThe main soft spot is not the reader's in-sample tuning objection, though that is valid. It is the evaluation metric. Equation (14) uses a 100x100 equidistant grid on Omega. For Pout, eps = 1e-8, so the outflow layers have width 1e-8; for Ppara, the parabolic layers have width 1e-4. No grid point lies in either layer. The reported discrete RMS error measures only the interior. A completely wrong layer contributes roughly 1e-4 to the true L2 error for Pout and roughly 1e-2 for Ppara; the paper claims errors around 1e-4 and 6e-5. So the central claim that these architectures lead to a tremendous increase of accuracy is not established for the defining feature of the problem. The hand-tailored indicators in (16) and (17), chosen with knowledge of where the layers are, further muddy the source of improvement.\n\nOther soft spots are less severe: hyperparameters are selected on the same two test problems, so the quantitative gains are partly fitted; key tables lack error bars; and there are only two benchmarks. These are fixable. The learned-tau mechanism is not tautological, since tau is trained from the residual rather than from the exact solution.\n\nWho benefits: researchers working on VPINNs and SciML for singularly perturbed PDEs. The learned-tau idea may prove useful even if the evidence here is not decisive. I would send it to peer review, with a request to fix the metric or supplement it with a layer-resolving error measure before publication.","headline":"Two sensible, incremental VPINN extensions, but the reported L2 errors are sampled on a grid that cannot see layers of width 1e-8 or 1e-4, so the claimed accuracy gains are not actually demonstrated.","tokens_in":16161,"tokens_out":2245,"would_cite":true,"duration_ms":25006,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N35","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding a network-learned SUPG stabilization term and an adaptive hard-constraint boundary indicator to variational physics-informed neural networks substantially improves accuracy on convection-dominated…","keywords":["convection-diffusion-reaction problems","convection-dominated problems","hp-VPINNs","FastVPINNs","SUPG stabilization","hard-constrained Dirichlet boundary conditions","adaptive indicator function","physics-informed neural networks"],"falsifier":"Re-run the two benchmark problems with all hyperparameters chosen before any solution information is used and with errors reported from a fixed run rather than the best epoch; if the learned-$\\tau$ and adaptive-indicator errors no longer beat the plain variational loss, the central claim is falsified. Alternatively, apply the same architecture to a problem with a curved or internal layer and check whether the per-boundary exponential indicator family keeps the error below the plain-loss level.","tokens_in":15163,"feed_emoji":"🎯","tokens_out":7961,"duration_ms":90723,"temperature":0.7,"pith_summary":"The paper sets out to make variational physics-informed neural networks (hp-VPINNs) accurate on convection-dominated convection-diffusion-reaction problems, where true solutions have thin boundary layers and standard loss functionals struggle. It proposes two extensions: a streamline-upwind Petrov–Galerkin (SUPG) stabilization term with a spatially varying stabilization parameter predicted by the network, and an adaptive indicator function whose boundary steepness exponents are learned during training. On benchmark problems with diffusion coefficient $\\varepsilon=10^{-8}$, the learned-stabilization version lowers the $L^2$ error from $1.693\\times10^{-4}$ to $1.037\\times10^{-4}$ for the outflow-layer problem and from $1.544\\times10^{-4}$ to $9.043\\times10^{-5}$ for the parabolic-layer problem; the adaptive indicator improves this further to $5.887\\times10^{-5}$ and $6.007\\times10^{-5}$, respectively. These values lie well below the errors reported for PINN and hp-VPINN baselines in the literature. If the result holds, it provides a practical route to resolving boundary layers without resolving their width on a mesh, and it removes two hand-tuning burdens in physics-informed neural networks.","feed_headline":"Learned stabilization and adaptive boundary cut PINN error to 6e-5","feed_subtitle":"On convection-dominated problems, a network-learned SUPG parameter and boundary indicator beat hand-tuned alternatives.","key_machinery":"The load-bearing object is the hard-constrained ansatz $u^{\\text{hard}}_{\\text{NN}}(x) = j(x) + h(x) u_{\\text{NN}}(x;\\theta)$, where $j$ extends the Dirichlet data, $h$ vanishes on $\\partial\\Omega$ and is positive inside, and boundary conditions are therefore satisfied exactly. On top of this, the SUPG term $L_{\\text{SUPG}} = \\int_\\Omega \\tau(x)(b\\cdot\\nabla u + cu - f)(b\\cdot\\nabla v)\\,dx$ adds streamline diffusion; the stabilization parameter is not a tuned constant but a second network output, $\\tau(x) = \\tau_g \\tanh(50x)\\tanh(50y)\\tanh(50(1-x))\\tanh(50(1-y))\\,\\sigma(\\tau_{\\text{NN}}(x;\\theta))$, which is positive, spatially varying, and zero at the boundary. For the indicator, the mechanism is an exponential family with per-boundary exponents $\\kappa_i = 10^{\\text{learnable exponent}}$, which lets the network adapt the slope of $h$ near each boundary to the steepness of the solution's layers. These two mechanisms together carry the reported accuracy improvement.","core_discovery":"The central claim is that two modifications to the hp-VPINN loss functional turn a method that historically struggles on convection-dominated problems into one that resolves boundary layers at $\\varepsilon=10^{-8}$ to roughly $10^{-4}$ accuracy on the studied benchmarks. First, replacing the standard variational loss $L^{\\text{hard}}_{\\text{var}}$ with $L^{\\text{hard}}_{\\text{var}} + L_{\\text{SUPG}}$, where $L_{\\text{SUPG}} = \\int_\\Omega \\tau(x)(b\\cdot\\nabla u + cu - f)(b\\cdot\\nabla v)\\,dx$, and having the network predict a positive, spatially varying $\\tau$ that vanishes on the boundary, consistently outperforms both the plain loss and the same loss with a constant, hand-tuned $\\tau$. Second, in the hard-constraint ansatz $u^{\\text{hard}}_{\\text{NN}} = j(x) + h(x)u_{\\text{NN}}$, letting $h$ be of the form $(1-e^{-\\kappa_1 x})(1-e^{-\\kappa_1 y})(1-e^{-\\kappa_2(1-x)})(1-e^{-\\kappa_2(1-y)})$ with $\\kappa_i = 10^{\\alpha}, 10^{\\beta}$, and learning $\\alpha,\\beta$ (and a third exponent for the parabolic-layer problem) alongside the solution, gives another accuracy gain and is stable across $\\varepsilon = 10^{-4}, 10^{-6}, 10^{-8}$. The paper reports these combined choices as producing a large increase in accuracy.","pith_inferences":["A natural extension is to compare the spatially varying $\\tau(x,y)$ learned by the network with a-posteriori optimized SUPG parameters; if the patterns match, it would support the view that the network discovers the same stabilization distribution as classical parameter optimization.","Because the learned indicator exponents varied only weakly with the Péclet number, one could pretrain the indicator branch on a cheap small-$\\varepsilon$ run and reuse it for nearby diffusion coefficients; the paper does not test this transfer.","The exponential indicator family used here assigns one exponent per boundary segment, so curved or interior layers would likely require a richer parameterization; the reported accuracy gains may not automatically transfer to non-axis-aligned layers.","The paper reports best-epoch and, in the constant-parameter tables, best-run errors after tuning hyperparameters against the known exact solution; a stricter protocol that withholds the exact solution during tuning would clarify the size of the practical gain."],"forward_implications":["On the two studied benchmarks, the combined method reaches $L^2$ errors below $1.1\\times10^{-4}$ for the outflow-layer problem and below $6.1\\times10^{-5}$ for the parabolic-layer problem at $\\varepsilon=10^{-8}$, below the reported PINN and hp-VPINN baselines.","A network-predicted stabilization parameter removes the need for a grid search over $\\tau$, and its spatial variation concentrates stabilization where the solution has large streamline derivatives.","Learning the indicator exponents provides a way to satisfy Dirichlet conditions exactly while adapting the near-boundary slope to layers, which the paper argues is essential for inlet and boundary layers.","The learned indicator parameters are nearly independent of $\\varepsilon$ across the range $10^{-4}$ to $10^{-8}$, suggesting the mechanism is capturing layer geometry rather than a particular diffusion value.","Adding only an $L^2$ weight-regularization term did not improve accuracy, indicating that the paper's route to accuracy is stabilization and boundary representation, not weight decay."],"supporting_citations":[{"why":"Supplies the two benchmark problems, the baseline PINN/hp-VPINN errors the paper compares against, and the standard residual-loss setup that motivates the new loss functionals.","marker":"Frerichs-Mihov et al. (2024a)"},{"why":"Introduces the multidimensional SUPG stabilization concept that is adapted into the continuous SUPG loss term.","marker":"Hughes and Brooks (1979)"},{"why":"Formulates the streamline-upwind Petrov–Galerkin method for convection-dominated flows, the classical basis for the stabilization term.","marker":"Brooks and Hughes (1982)"},{"why":"Shows that optimized SUPG stabilization parameters can reduce spurious oscillations, motivating a spatially varying network-predicted $\\tau$.","marker":"John et al. (2011)"},{"why":"Provides the hard-constraint ansatz $u = j(x) + h(x)u_{\\text{NN}}$ used to impose Dirichlet boundary conditions exactly.","marker":"Lu et al. (2021b)"},{"why":"Defines the hp-VPINN test-function framework, Legendre polynomial test functions, and Gauss–Lobatto–Legendre quadrature used in the implementation.","marker":"Kharazmi et al. (2021)"},{"why":"Supplies the Eriksson–Johnson benchmark and comparison data used to validate the FastVPINNs implementation.","marker":"Sikora et al. (2023)"},{"why":"Discusses the difficulty of choosing the indicator function for hard-constrained boundary conditions, motivating the adaptive indicator approach.","marker":"Matthaiou et al. (2024)"},{"why":"Provides the algebraically stabilized MUAS method used to generate the reference solution for the parabolic-layer problem.","marker":"John and Knobloch (2022)"}],"fun_headline_variants":["Learned SUPG and adaptive boundary cut PINN error","Network-learned stabilization and boundary gate boost PINN","Adaptive boundary and learned SUPG beat hand-tuned PINNs","Self-predicted SUPG and boundary improve VPINN accuracy","PINN learns its own stabilization and boundary indicators"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The improvements are measured with hyperparameters (constant $\\tau$, $\\tau_g$, and indicator exponents) selected by searching against the known exact solution and reporting the best epoch and best run, so if those hyperparameters had to be fixed without knowing the solution the reported gains could shrink or vanish; a second assumption is that the per-boundary exponential indicator family can represent the layer geometry of general convection-dominated solutions.","fun_headline_variants_meta":{"raw":{"variants":["Learned SUPG and adaptive boundary cut PINN error","Network-learned stabilization and boundary gate boost PINN","Adaptive boundary and learned SUPG beat hand-tuned PINNs","Self-predicted SUPG and boundary improve VPINN accuracy","PINN learns its own stabilization and boundary indicators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000431,"raw_usage":{"total_tokens":2223,"prompt_tokens":991,"completion_tokens":1232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1150}},"tokens_in":607,"tokens_out":1232,"duration_ms":13255,"temperature":1.0,"reasoning_tokens":1150,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:45:47.497096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the two benchmark problems with all hyperparameters chosen before any solution information is used and with errors reported from a fixed run rather than the best epoch; if the learned-$\\tau$ and adaptive-indicator errors no longer beat the plain variational loss, the central claim is falsified. Alternatively, apply the same architecture to a problem with a curved or internal layer and check whether the per-boundary exponential indicator family keeps the error below the plain-loss level.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that optimized SUPG stabilization parameters can reduce spurious oscillations, motivating a spatially varying network-predicted $\\tau$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the hp-VPINN test-function framework, Legendre polynomial test functions, and Gauss–Lobatto–Legendre quadrature used in the implementation."},{"cited_title":"Physics informed neural networks with strong and weak residuals for advection-dominated diffusion problems, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the Eriksson–Johnson benchmark and comparison data used to validate the FastVPINNs implementation."},{"cited_title":"Bound-preserving PINN s for steady-state convection-diffusion-reaction problems","cited_arxiv_id":null,"evidence_quote":"Discusses the difficulty of choosing the indicator function for hard-constrained boundary conditions, motivating the adaptive indicator approach."},{"cited_title":"On algebraically stabilized schemes for convection-diffusion-reaction problems","cited_arxiv_id":null,"evidence_quote":"Provides the algebraically stabilized MUAS method used to generate the reference solution for the parabolic-layer problem."}],"review_version":1}