{"id":"c1c6391f-efc7-4baf-80c3-e9a41b10ca24","arxiv_id":"2412.00579","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A tensorized Fourier neural operator trained with an H2 loss predicts macroscopic permeability from Stokes-Brinkman coefficient fields with about 9% relative error on log-permeability.","lead":"The authors train Fourier neural operators to predict, from a heterogeneous coefficient field, the velocity field and then the macroscopic permeability of flow in dual-scale porous media. They compare seven loss regularizations and find that penalizing second derivatives of the velocity gives the most accurate permeability estimates on synthetic Stokes-Brinkman problems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Local permeability standard deviation of 1e-12 in data generation collapses beta to a binary mask, so the reported errors may not generalize to genuinely heterogeneous dual-scale fabrics.","rationale":"The reader's weakest assumption is exactly the 1e-12 standard deviation issue, and I agree that this is the most load-bearing concern. The paper's empirical claims are internally consistent on the dataset as generated, but the data generation undermines the stated scope: the introduction and abstract emphasize 'extreme heterogeneity (spanning several orders of magnitude)' in beta, yet a standard deviation of 1e-12 in log(ks) makes the local permeability field essentially uniform within each sample. The central result, that H2 regularization improves permeability prediction, is plausible only for the binary-mask data; it is not evidence for the more general problem where beta varies continuously over orders of magnitude within the fibre bundles. This is not an attack on the authors' integrity, but a concrete mismatch between the claimed problem class and the data used to support it. The issue is addressable: a revision that generates data with realistic within-sample permeability variation, or clearly limits the claims to two-phase constant-solid-permeability media, would resolve it. Therefore the CONDITIONAL verdict stands, and no change to the reader's judgment is needed. The proposed concrete test directly checks whether the simplification matters by comparing performance on data with larger internal variation against the reported error.","tokens_in":11402,"tokens_out":5715,"duration_ms":56787,"concrete_test":"Generate a new held-out test set with the same mean, anisotropy, and rotation parameters, but draw log(ks) with a standard deviation of order 1.0 (or uniformly spanning 1e-11 to 1e-4 mm^2 within each field). Evaluate the already-trained TFNO H2 model on this set without retraining, or retrain with identical hyperparameters on the new distribution if zero-shot transfer is not expected. If the relative log-permeability error degrades from 0.0913 to above roughly 0.3, or if the ranking of loss functions changes materially, the std=1e-12 dataset simplification is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 states: 'We generate one sample for each set of parameters with the standard deviation fixed to a small value (10−12).' Since log(ks) is drawn from a Gaussian random field with this standard deviation, the generated local permeability is essentially constant within each sample. The mu/ks map is then nearly uniform, and thresholding at the 80% quantile produces a binary mask: beta is 0 in ~20% of pixels and roughly constant (equal to mu/mean(ks)) in the rest. The only magnitude variation of beta arises from the mean parameter, which changes across the 800 parameter combinations, not within a single realization. The paper repeatedly emphasizes 'heterogeneity spanning several orders of magnitude' in the governing equation, but that heterogeneity exists only across the dataset, not inside a field. Consequently, the operator learning task reduces to geometry segmentation with a single solid-phase beta value, and the reported TFNO H2 log-permeability error of 0.0913 may not transfer to real textile fabrics where local permeability varies smoothly over orders of magnitude within fibre bundles. The paper does not discuss this consequence or validate the method on fields with internal permeability variation, so the central claim about solving the 'challenging dual-scale flow problem' is not yet supported for the stated problem class.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-stage surrogate for the Stokes–Brinkman dual-scale flow problem: a Fourier neural operator (FNO/TFNO) maps the heterogeneous coefficient β to a velocity field, from which the macroscopic permeability K is recovered by volume averaging and Darcy upscaling. Using 800 generated samples, the authors compare seven loss formulations (L2, H1, βu-regularized H1, βu-plus-Laplacian H1, H2, βu-regularized H2, and macroscopic pressure-gradient). On the test set, the TFNO trained with the H2 loss achieves the lowest reported log-permeability error (J = 0.0913) and retains accuracy in zero-shot super-resolution from 64×64 to 128×128 grids. The central claim is that regularizing second derivatives of the velocity during training significantly improves permeability prediction.","tokens_in":11561,"tokens_out":6619,"duration_ms":66464,"significance":"If the result holds, the paper would be a useful step toward cheap permeability surrogates for liquid composite moulding, and the H2-regularization insight is physically plausible because computing K involves the Laplacian of the velocity field. The zero-shot super-resolution result is practically attractive, and the empirical comparison in Tables 1 and 2 is internally consistent. However, the significance is currently limited by three issues: the input fields are effectively binary due to the 1e-12 GRF standard deviation, no error bars or seed statistics are given, and several loss-function hyperparameters are unreported. These gaps make it difficult to assess whether the method transfers to genuinely heterogeneous dual-scale fabrics and whether the reported rankings are robust.","major_comments":[{"comment":"The choice of Gaussian random field standard deviation fixed to 10^{-12} makes log(k_s) essentially constant within each generated realization, so the thresholded input β is a binary mask with a single solid-phase value. The manuscript repeatedly emphasizes 'extreme heterogeneity spanning several orders of magnitude', but that heterogeneity is present only across the 800 samples (through the mean parameter), not inside any field, and the operator learning task reduces to geometry segmentation plus a scalar magnitude. The central claim that the method addresses the challenging dual-scale flow problem with large local permeability variation is therefore not supported for the stated problem class. Please either justify the 1e-12 choice, generate fields with internal log-permeability variation spanning several orders of magnitude within a single realization, and report errors on such fields, or explicitly reframe the contribution as a two-phase geometry problem.","section":"§3 (Data generation)"},{"comment":"All reported errors come from a single training run for each loss/architecture combination, with no seed averaging, error bars, or significance testing. Several differences that drive the central conclusion are small (e.g., TFNO H2 J=0.0913 versus H1βuΔu J=0.0979 versus L2 J=0.1018 in Table 2), so the ranking may be within training noise. The claim that regularising second derivatives 'significantly enhances' the results needs repeated-seed statistics (mean ± standard deviation over at least three seeds) or a paired test across the fixed test set.","section":"§5, Tables 1 and 2"},{"comment":"The metric J := ∥ log(K_pred) − log(K_GT) ∥ / ∥ log(K_GT) ∥ with ∥K∥ = max_{i,j} |k_ij| is not well defined if any component of the permeability tensor is negative, and a logarithm of a matrix/tensor is not standard. For rotated anisotropic media the off-diagonal components of K can be negative, so the paper must specify the exact elementwise transformation (e.g., log of absolute values, log of eigenvalues, or another positive quantity) and confirm that it is defined on all 200 test samples.","section":"§5.2, Table 2"},{"comment":"The regularization weights λ, λ1, and λ2 are introduced in the loss definitions but their values (or tuning procedure) are never reported. Since the central comparison is between different regularizations, the absence of these hyperparameters makes the experiments irreproducible and prevents the reader from separating the effect of the loss form from the effect of the weight choice.","section":"§4, Eqs. (15), (16), (19)"},{"comment":"The neural operator input is described as log(β) or log10(β), while β = 0 in the clear-fluid regions. The paper does not state how the zero values are encoded (e.g., a large negative sentinel, a mask channel, beta+epsilon), which is essential for reproducing the model and for interpreting Figures 3–4.","section":"§3–§4"}],"minor_comments":[{"comment":"There are several typos: 'Since we our primary motivation' should be 'Since our primary motivation', 'corse grained' should be 'coarse grained', and 'surpremum' should be 'supremum'.","section":"§5.2, §5.3, Table 2"},{"comment":"The sentence 'We generate one sample for each set of parameters' implies that the GRF randomness is effectively unused; please clarify how the 80% quantile thresholding interacts with the near-constant field and how the mask geometry is controlled.","section":"§3"},{"comment":"The visual comparison is based on a single test/training sample per model; please state whether the displayed examples are representative and/or provide quantitative error maps for the full test set.","section":"§5.1, Figures 3 and 4"},{"comment":"The mixed-derivative term in the definition of the H2 norm appears to be missing the ∂² symbol in the typeset formula; please correct the notation so that all second partial derivatives are explicitly shown.","section":"§4, Eq. (18)"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely within scope for a fluids/ML-oriented journal, and the core two-stage idea is reasonable. The data-generation concern is the main load-bearing issue: with a GRF standard deviation of 1e-12, the experiments do not exercise the claimed heterogeneous-permeability regime. I would not reject outright, but the authors need to either add experiments with genuinely varying local permeability or carefully reframe the contribution. The missing hyperparameter values and the undefined metric also need to be fixed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent empirical study: it trains FNO and TFNO on the Stokes-Brinkman beta-to-velocity map, compares seven loss functions, and finds that TFNO with an H2 loss gives the best log-permeability error (0.0913). The application to this dual-scale flow problem is new, and the systematic comparison of regularizations is a practical contribution for the liquid composite molding community. The two-step approach (predict velocity, then upscale to K) is sensible, and the zero-shot super-resolution test is a nice addition.\n\nThe main soft spot is the data generation. Section 3 fixes the GRF standard deviation to 1e-12, which makes log(ks) essentially constant within each sample. The resulting beta field is a binary mask with a single solid-phase value; the 'heterogeneity spanning orders of magnitude' only exists between samples (via the mean), not inside a field. So the test set covers a much narrower problem class than the paper claims. This does not invalidate the loss comparison for binary geometries, but it does mean the reported errors may not transfer to real fabrics with smoothly varying local permeability. The stress-test note is right about this.\n\nOther issues are minor: no error bars or seed averages, so the loss ranking is not robust to stochasticity; the explanation that H2 regularization matters because K involves the Laplacian is misleading, since the volume-averaged Laplacian vanishes under periodic boundary conditions (the benefit is indirect through better velocity accuracy); and no code/data/hyperparameters are provided, which limits reproducibility.\n\nOn balance, the empirical claims are internally consistent and the loss comparison is a useful benchmark for practitioners. The data-generation issue is addressable and should not be fatal to the paper's central contribution, provided the authors either fix the data or narrow the claim. I'd send it to peer review: a competent referee can help sharpen the framing and ask for error bars and code. For a reading group, it's worth a look if you're working on operator learning for porous media; otherwise it's a nice-to-have.","headline":"Useful loss-function benchmark for neural-operator permeability prediction, but the 1e-12 GRF standard deviation makes the test fields effectively binary, so the 'heterogeneous local permeability' claim is overstated.","tokens_in":12184,"tokens_out":4436,"would_cite":false,"duration_ms":44567,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["47.56.+r"],"model":"deepseek-v4-flash","headline":"A neural operator trained with H2 regularization maps the Stokes-Brinkman coefficient to a velocity field and recovers macroscopic permeability with a relative log-error of about 0.09.","keywords":["Dual-scale flow","Porous media","Neural operator","Composite manufacturing","Permeability","Stokes-Brinkman equation","Fourier neural operator","Loss regularization"],"falsifier":"Run the same TFNO-$H^2$ training on a dataset whose local permeability maps have standard deviation varying over several orders of magnitude rather than fixed at $10^{-12}$, and compare the relative $\\log K$ error on the test set. If the error rises substantially above $0.0913$, the claim that the method handles heterogeneous dual-scale fabrics is not supported.","tokens_in":11103,"feed_emoji":"💧","tokens_out":9195,"duration_ms":128736,"temperature":0.7,"pith_summary":"The paper replaces the expensive Stokes-Brinkman solver with a neural operator that maps the heterogeneous coefficient $\\beta$ to the velocity field, then recovers the macroscopic permeability tensor by upscaling. It claims the choice of loss is decisive: penalizing the second derivatives of the predicted velocity field during training, through an $H^2$ loss or an $H^1$ loss with additional $\\beta u$ and Laplacian terms, markedly improves the computed permeability. The best Tensorized Fourier Neural Operator achieves a relative log-permeability error of $0.0913$ on the test set and degrades only slightly under zero-shot super-resolution from $64\\times64$ to $128\\times128$. A fast, accurate surrogate for permeability matters because direct numerical simulation of dual-scale fibrous media is computationally prohibitive for industrial process design.","feed_headline":"Neural operator predicts dual-scale fabric permeability to 9%","feed_subtitle":"Second-derivative regularization makes a Stokes-Brinkman surrogate accurate enough for composite moulding process design.","key_machinery":"The central objects are the Stokes-Brinkman equation $\\varphi\\Delta u - \\beta u - \\nabla p = 0$ with periodic conditions, and a two-step surrogate: a Fourier neural operator (a network whose integral kernel acts by multiplication in Fourier space, here the Tensorized Fourier Neural Operator with Tucker factorization) that approximates the map $\\beta \\mapsto u$, followed by Darcy upscaling $U = -K G/\\mu$ with $G = \\langle \\varphi\\Delta u \\rangle_{\\Omega} - \\langle \\beta u \\rangle_{\\Omega}$. The loss functions carry the argument: relative $L^2$, $H^1$, $H^2$, and variants regularized by $\\beta u$ and $\\Delta u$ terms, with $H^2$ and $\\beta u\\Delta u$-regularized $H^1$ performing best for permeability. The model's resolution independence is what allows zero-shot super-resolution tests from a $64\\times64$ to a $128\\times128$ grid.","core_discovery":"The central claim is that regularizing the second derivatives of the velocity field is what makes operator learning viable for the dual-scale permeability problem. Directly learning the map $\\beta \\mapsto K$ proved unstable because it requires inverting the macroscopic pressure gradient, and an unregularized $L^2$ fit to the velocity field yields poor permeability values. The paper establishes that the $H^2$ loss, which explicitly penalizes differences in the second partial derivatives of predicted and ground-truth velocity, gives the best macroscopic permeability, with a relative supremum-norm error $J = \\|\\log K_{\\text{pred}} - \\log K_{\\text{GT}}\\|/\\|\\log K_{\\text{GT}}\\|$ of $0.0913$ for the TFNO model. The authors conclude that this highlights the importance of second-derivative regularization because computing $K$ involves applying the Laplacian to the velocity field.","pith_inferences":["Inference: the 800-sample training set, with the standard deviation of $\\log(k_s)$ fixed at $10^{-12}$, effectively makes each local permeability field nearly constant, so the reported error is a benchmark on near-binary $\\beta$ fields rather than on fields with local permeability varying over many orders of magnitude.","Inference: because the best losses explicitly target $\\Delta u$, the same regularization principle could transfer to operator learning for other elliptic problems where the quantity of interest is a functional of the solution's second derivatives, such as stress or dissipation.","Inference: a natural extension is to train with the standard deviation of the Gaussian random field sampled over several orders of magnitude, which would test whether the $H^2$ advantage persists when local permeability genuinely varies across the domain."],"forward_implications":["The TFNO-$H^2$ surrogate computes macroscopic permeability at near-zero marginal cost, making it usable in design loops where thousands of Stokes-Brinkman solves would be too slow.","Because resolution independence persists, a model trained on coarse grids can be evaluated on finer microstructural images, easing transfer from simulation grids to high-resolution fabric scans.","The result identifies second-derivative fidelity, not just pointwise velocity accuracy, as the quantity neural operators should preserve when the observable of interest involves the Laplacian of the field.","The same two-step strategy (predict the field with a neural operator, then apply the physical upscaling operator) is directly applicable to other upscaling problems governed by PDEs with heterogeneous coefficients."],"supporting_citations":[{"why":"Supplies the Anderson-accelerated FFT solver that generates the ground-truth velocity fields and permeabilities.","marker":"[8]"},{"why":"Defines the neural operator framework whose lifting, Fourier layer, and projection structure the models use.","marker":"[11]"},{"why":"Provides the parameter-to-observable-map perspective that motivates predicting permeability as a quantity of interest.","marker":"[12]"},{"why":"Gives the FFT moving average method for sampling the Gaussian random field inputs.","marker":"[15]"},{"why":"Introduces the Fourier Neural Operator whose spectral convolution layer is the backbone of both models.","marker":"[31]"},{"why":"Introduces the Tensorized Fourier Neural Operator with Tucker factorization that achieves the best permeability errors.","marker":"[37]"}],"fun_headline_variants":["Second-derivative penalty boosts neural operator permeability accuracy","H2 loss sharpens Stokes-Brinkman operator learning for composites","Regularizing velocity curvature yields 9% permeability error","Why second-derivative regularization matters in dual-scale flow","Neural operator permeability prediction improved by Hessian penalty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim assumes the Gaussian random field samples, with the standard deviation of $\\log(k_s)$ fixed at $10^{-12}$, are representative of the heterogeneous dual-scale porous media the model is meant to predict.","fun_headline_variants_meta":{"raw":{"variants":["Second-derivative penalty boosts neural operator permeability accuracy","H2 loss sharpens Stokes-Brinkman operator learning for composites","Regularizing velocity curvature yields 9% permeability error","Why second-derivative regularization matters in dual-scale flow","Neural operator permeability prediction improved by Hessian penalty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1160,"prompt_tokens":854,"completion_tokens":306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":228}},"tokens_in":470,"tokens_out":306,"duration_ms":63614,"temperature":1.0,"reasoning_tokens":228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:12:29.681723+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same TFNO-$H^2$ training on a dataset whose local permeability maps have standard deviation varying over several orders of magnitude rather than fixed at $10^{-12}$, and compare the relative $\\log K$ error on the test set. If the error rises substantially above $0.0913$, the claim that the method handles heterogeneous dual-scale fabrics is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Anderson-accelerated FFT solver that generates the ground-truth velocity fields and permeabilities."},{"cited_title":"Kovachki, Z","cited_arxiv_id":null,"evidence_quote":"Defines the neural operator framework whose lifting, Fourier layer, and projection structure the models use."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the parameter-to-observable-map perspective that motivates predicting permeability as a quantity of interest."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the FFT moving average method for sampling the Gaussian random field inputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Fourier Neural Operator whose spectral convolution layer is the backbone of both models."}],"review_version":1}