{"id":"13dab8c3-d67d-4608-a5b7-ed688787e8ff","arxiv_id":"2502.05044","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-scale physics-informed neural network, regularized by a mesoscale flow solver, predicts fibrous tow permeability more accurately than standalone PINNs in 2D benchmarks, while simpler upscaling methods stay close to fully resolved simulations at much lower cost.","lead":"This paper tests four ways to predict how easily resin flows through woven fiber textiles, from simple upscaling to full 3D simulation, and adds a two-scale physics-informed neural network solver. It matters because composite manufacturing needs fast and reliable permeability estimates, and the hybrid solver shows promise on small 2D problems while the full-resolution reference model costs far more compute.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hybrid solver's coarse-scale regularizer is self-referential: the Stokes-Brinkman target in Eq. (14) is built from the PINN's own projected permeability, so reported accuracy is not shown to be independent of the hand-set bounds [5e-5,5e-4] and initial guess 4.5e-4.","rationale":"The central claim rests on the hybrid solver's ability to use mesoscale information to regularize the microscale PINN. For this to be a genuine improvement, the mesoscale target in R(theta, psi) must constrain the PINN in a way that is not already determined by the PINN's own output. But in Algorithm 1, the permeability dKp11 used to build the Stokes-Brinkman target is computed from the current PINN state and then projected onto a hand-set interval. The projection (18) is the only mechanism preventing wildly wrong dKp values from entering the mesoscale solve. Since the reference permeabilities for both benchmark geometries lie inside [5e-5,5e-4], the interval encodes substantial prior knowledge of the answer. Without a sensitivity analysis, the reported convergence could be a consequence of this prior rather than of the coupling. The reader's weakest_assumption identifies this correctly; I agree, with additional emphasis that this is a feedback loop, so the regularizer is not an independent validation signal. To its credit, the paper does report detailed hyperparameters and honestly notes the SBM-vs-NUM negative result and the limitations of the 2D benchmarks. The engineering comparison has a separate baseline mismatch and an internal runtime inconsistency, but these do not determine the hybrid claim. The appropriate verdict remains conditional: the idea is plausible and testable, but the paper must supply an ablation before the central claim can be accepted.","tokens_in":15079,"tokens_out":10349,"duration_ms":105338,"concrete_test":"Rerun Algorithm 1 on VZ1 and VZ2 with a sensitivity sweep: (a) no projection, (b) Kp_LB and Kp_UB varied over intervals such as [1e-6,1e-2], [1e-4,3e-4] for V1, and [3e-5,2e-4] for V2, and (c) gKp in {1e-5,1e-4,4.5e-4,1e-3,1e-2}, keeping all other settings fixed. Record final l2 velocity and pressure errors and the final dKp11. If these remain close to the values in Fig. 8 for all configurations that still bracket the reference value, the concern is resolved; if convergence is lost when the guess or bounds move away from the reference, the reported result depends on prior knowledge of the answer rather than on the coupling mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 lines 8-10 compute dKp11 from the current PINN solution, project it via (18) onto the user-specified interval [Kp_LB,Kp_UB], then solve the Stokes-Brinkman problem (15) with that projected value and use the resulting velocity and pressure as the target in the coarse-scale regularizer R (14). Thus the mesoscale information that rescues the PINN is not an independent reference solution; it is a self-consistent reconstruction in which the same dKp appears both as the PINN's output and as the parameter that generates the training target. For the two reported geometries, the chosen interval [5e-5,5e-4] contains both reference permeabilities (2.37e-4 for V1 and 9.08e-5 for V2), and the initial guess 4.5e-4 is within a factor of about two of the V1 reference. The paper provides no ablation varying Kp_LB, Kp_UB, or gKp, and no analysis of whether other fixed points of this feedback loop exist. Consequently, the headline result, that coarse-scale regularization fixes PINN failure, may be an artifact of a privileged initialization and projection interval rather than a property of the dual-scale coupling. The engineering comparison in Section 6.1 is additionally weakened by the mismatch between NUM/SBM and FRM fiber volume contents in Table 1 and by the abstract's 45-minute runtime claim contradicting the 9 to 10.5 hour runtimes in the same table, although this is secondary to the hybrid claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a multiscale permeability-prediction framework for fibrous textiles, comparing four approaches: numerical upscaling (NUM), scale-bridging with machine-learning emulators (SBM), fully resolved models (FRM), and a newly developed hybrid dual-scale solver that couples a PINN microscale solver with a Stokes-Brinkman finite-element mesoscale solver. The central scientific claim is that the hybrid solver, through the coarse-scale regularizer R in Eq. (14), achieves reasonable accuracy within 25k iterations on two 2D periodic fiber geometries, whereas standard PINNs fail. A secondary engineering claim is that NUM and SBM reproduce FRM in-plane permeabilities while requiring far less runtime than FRM.","tokens_in":15462,"tokens_out":7028,"duration_ms":72290,"significance":"If the hybrid claim is established, the paper would make a useful contribution by showing that coarse-scale information can regularize PINN training under periodic boundary conditions, a known failure mode. The manuscript has genuine strengths: two well-defined 2D benchmarks with finite-element reference solutions, a detailed statement of Algorithm 1, explicit hyperparameter values, and unusually candid discussion of limitations in Sections 6.2 and 7. However, the current evidence does not yet establish that the coarse-scale regularizer is the cause of the improvement, because the regularizer depends on the very permeability that the PINN is learning, and the projection interval and initial guess are hand-chosen. The engineering comparison is also weakened by unmatched fiber volume contents and hardware differences. The paper is honest about several of these concerns but does not resolve them.","major_comments":[{"comment":"The coarse-scale regularizer R is not an independent reference. It evaluates the PINN velocity and pressure against the finite-element solution of (15), but that finite-element solution is computed using dKp11[VZ] extracted from the current PINN output and then projected by (18). The same quantity the PINN is learning therefore appears on both sides of the training signal, and the projection interval [5e-5, 5e-4] and initial guess 4.5e-4 are chosen so that both reference values (2.37e-4 for VZ1 and 9.08e-5 for VZ2) lie inside the interval. The paper provides no sensitivity study varying gKp, Kp_LB, or Kp_UB, and no analysis of other possible fixed points of this feedback loop. Without such an ablation, or a comparison against a regularizer built from a fixed reference permeability, the headline conclusion that coarse-scale regularization fixes PINN failure is not established; the reported accuracy may reflect favorable manual choices rather than a property of the dual-scale coupling.","section":"§5.3, Eq. (14), Algorithm 1 lines 8–10"},{"comment":"The accuracy numbers for the hybrid solver are single trajectories with no seed-to-seed variation or confidence intervals, and the network and optimization hyperparameters appear to be tuned on the same two benchmark geometries used for evaluation. The permeability history for VZ2 in Fig. 8(f) visibly overestimates the reference and oscillates, so the statement that the hybrid solver 'yields reasonable accuracy within 25k iterations' needs support from repeated runs with different initializations and reporting of mean and spread. Ideally the method should also be evaluated on a third geometry not used for hyperparameter selection. This is load-bearing because it is the main evidence for the claimed improvement over standard PINNs.","section":"§6.2, Fig. 8"},{"comment":"The claimed agreement between NUM/SBM and FRM is not quantified at matched conditions. The two families of models are reported at different fiber volume contents (NUM/SBM: 40.8%, 45.2%, 49.9%, 55.0%; FRM: 38.4%, 42.7%, 48.0%, 49.9%, 54.9%, 60.0%), on different model sizes, and on different hardware, and the paper itself warns that the FRM role as reference must be critically examined. As one example, NUM20 at 49.9% FVC reports K11 = 5.76e-11 m² while FRM30 at 48.0% FVC reports K11 = 1.23e-10 m², a factor-of-two difference that is difficult to reconcile with the statement of 'good consistency' except at the lowest FVC. Please provide an interpolation to common FVC values or otherwise quantify the comparison before claiming that NUM and SBM reproduce FRM in-plane permeabilities.","section":"§6.1, Table 1"},{"comment":"The abstract states that the scale-bridging method has 'modeling runtimes of ~45 minutes per simulation', while Section 6.1 states that NUM and SBM complete simulations in under 11 hours, Table 1 lists 9.0–10.5 hours, and the conclusion repeats 'below 11 hours'. The 45-minute figure is inconsistent with the body of the paper and must be corrected or explicitly explained. In addition, the abstract at the start of the submission uses the taxonomy SSM/SUM/SBM/FRM, whereas the abstract in the body uses NUM/FRM/SBM/PINN; the two should be aligned so that the reader knows which methods are actually compared.","section":"Abstract vs. §6.1 and Table 1"}],"minor_comments":[{"comment":"The text contains a typo: 'caluclated' should be 'calculated'.","section":"§4"},{"comment":"The notation 'λSB_i > 0' is not defined; the weights appearing in Eq. (14) are λu and λp.","section":"§5.3, after Eq. (14)"},{"comment":"Notation alternates between V1/VZ1, Kp/Kp11, and dKp11/cKp without consistent subscripts; please standardize notation for the two-scale models and the permeability symbols.","section":"§5.1–§5.3"},{"comment":"The caption uses '25f' and '36f' while the text refers to VZ1 and VZ2; define these abbreviations in the caption or use the same names as the text.","section":"Fig. 8 caption"},{"comment":"The column header 'R T in h' is unclear; spell out 'RT' as 'runtime' and separate it from the preceding column.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The hybrid-solver claim is interesting and worth publishing if the authors can provide the requested ablations and repeated-run statistics. I would not accept the paper in its current form because the self-referential regularizer and the absence of sensitivity analysis leave the central claim under-supported. The manuscript fits the journal's scope, and the authors' candid limitation statements are a strength that should be retained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the hybrid PINN dual-scale solver is a real idea, but the paper does not yet show that it works for reasons independent of good hand-tuning. The stress-test note is on target.\n\nWhat's new: coupling a microscale PINN to a mesoscale Stokes-Brinkman FEM solver via a coarse-scale regularizer, applied to fibrous tow permeability. On two 2D periodic fiber geometries, standard PINNs fail to converge, while the hybrid version gets l2 velocity errors around 1e-2 and pressure errors around 1e-1, and for V1 recovers permeability close to the FE reference. That is a genuine demonstration, preliminary but not trivial, that mesoscale info can stabilize PINN training. The engineering comparison of NUM vs SBM vs FRM is also useful: SBM does not beat NUM, FRM is 88-120 h on a cluster vs 9-10.5 h for NUM/SBM on a workstation. The negative SBM result is worth publishing.\n\nSoft spots, in order of severity. First, Eq. (14) uses a Stokes-Brinkman target built from the PINN's own projected permeability. The projection interval [5e-5,5e-4] contains both reference values, and the initial guess 4.5e-4 is close to V1's reference. With no ablation on the bounds or initial guess, the reported improvement could be partly the hand-picked prior, not the dual-scale coupling. This is the load-bearing gap. Second, hyperparameters are tuned on the same two benchmarks; no seeds, no confidence intervals. VZ2's permeability is overestimated and unstable. Third, the abstract says ~45 minutes per simulation while the full text and Table 1 say 9-10.5 hours. That discrepancy must be fixed before publication. Fourth, no code or data is provided, so nothing here is independently reproducible.\n\nThe paper is honest about being a first step for 2D only, and the writing is clear. Who it's for: composites permeability modelers and PINN-for-PDE researchers. It deserves a serious referee, not a desk reject, but the referee should demand sensitivity analysis, code/data, and a reconciled runtime claim.","headline":"Hybrid PINN dual-scale solver is a plausible idea but the self-referential regularizer and a runtime contradiction in the abstract weaken the main claims; still worth serious review.","tokens_in":16016,"tokens_out":3272,"would_cite":false,"duration_ms":32456,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid dual-scale solver that couples a microscale physics-informed neural network (PINN) to a mesoscale Stokes-Brinkman finite element solver predicts permeability where standard PINNs fail.","keywords":["permeability prediction","scale bridging","physics-informed neural networks","Stokes-Brinkman equation","fibrous composites","multiscale simulation","periodic boundary conditions","textile permeability"],"falsifier":"Run the hybrid solver on the 25-fiber geometry with the projection upper bound set below the finite element reference value $2.37\\times10^{-4}$, for instance $K^p_{\\mathrm{UB}}=2.0\\times10^{-4}$; if the predicted permeability lands on the upper bound, then the hand-chosen interval, not the learned flow, is setting the reported accuracy.","tokens_in":14835,"feed_emoji":"🧵","tokens_out":18581,"duration_ms":158667,"temperature":0.7,"pith_summary":"This paper proposes a scale-bridging framework for predicting the permeability of fibrous textile structures and compares four routes: a fully resolved model (FRM), a numerical upscaling method (NUM), a data-driven scale-bridging method (SBM), and a hybrid dual-scale solver built on a physics-informed neural network (PINN). The paper claims that NUM and SBM reproduce the in-plane permeabilities of the expensive FRM reference while running in under 11 hours instead of 88 to 120 hours, although SBM does not clearly beat NUM. The central physics-informed claim is that standard PINNs fail on 2D periodic fiber geometries, but coupling the microscale PINN to a mesoscale Stokes-Brinkman finite element solver through a coarse-scale regularizer yields reasonable accuracy within 25,000 iterations. If correct, this shows that coarse-scale information can regularize PINN training under periodic boundary conditions, and it makes fast upscaling practical for composite manufacturing.","feed_headline":"Two-scale neural net cracks fiber flow single-scale nets miss","feed_subtitle":"Coupling two scales lets the network learn periodic fiber flows and predict permeability in minutes instead of days.","key_machinery":"The load-bearing mechanism is the coarse-scale regularizer $R(\\theta,\\psi)$ in Eq. (14), a coupling term added to the PINN objective that penalizes, at sampled mesh points inside the fluid region, the squared difference between the microscale PINN velocity and pressure and the velocity and pressure of a Stokes-Brinkman finite element solution on the mesoscale. Around it sits the iterative loop of Algorithm 1: the mesoscale problem (15) uses a permeability computed from the current PINN output through Darcy's law, that permeability is projected into $[K^p_{\\mathrm{LB}}, K^p_{\\mathrm{UB}}]$ by Eq. (18), the coupling term is refreshed every $T$ iterations, and after $k_c$ iterations the coupling weights are annealed so the microscale loss takes over. Periodic boundary conditions are encoded in the network architecture through periodic embeddings, while Fourier features and gradient-based weight scaling address spectral bias and stiff gradient flow.","core_discovery":"Standard physics-informed neural networks (PINNs) with periodic boundary conditions, Fourier features, and gradient-based loss balancing stagnate on the microscale Stokes problem in fiber perforations, with errors making no progress after 100,000 iterations; the paper attributes this to the limited information contained in periodic conditions and to stiff gradient flow from the no-slip penalty. The proposed hybrid dual-scale solver repairs this by adding the coarse-scale regularizer $R(\\theta,\\psi)$ from Eq. (14), which penalizes the distance between the microscale PINN velocity and pressure and a Taylor-Hood finite element solution of the mesoscale Stokes-Brinkman problem (15). The permeability predicted from the PINN output is projected into $[5\\times10^{-5},\\,5\\times10^{-4}]$ by Eq. (18), and the coupled objective is updated every $T=250$ iterations. Within 25,000 iterations the hybrid solver reaches $\\ell^2$ errors around $10^{-2}$ to $10^{-1}$ against the finite element reference and reproduces the reference permeability for the 25-fiber geometry inside the reported error band; for the 36-fiber geometry the flow is captured but the predicted permeability overestimates the reference and is less stable. The same paper reports that NUM and SBM reproduce FRM in-plane permeabilities with runtimes under 11 hours, while out-of-plane $K_{33}$ remains much more variable.","pith_inferences":["A testable extension is to shift the projection interval or initial guess and watch the predicted permeability; if the answer tracks the clip, the bounds rather than the physics carry the result.","The same coarse-scale-regularization idea could rescue PINNs in other multiscale PDEs where periodic boundary conditions underdetermine the fine-scale solution, by coupling the fine network to a homogenized or coarse-grid solver.","The contrast between the stable 25-fiber case and the overestimating 36-fiber case suggests a stiffness threshold as fiber count grows; measuring error versus fiber number would map where the regularizer stops compensating.","For manufacturing practice, the results imply NUM with uniform micropermeability may be sufficient for in-plane permeability, so the extra complexity of segment-wise ML surrogates should be reserved for out-of-plane or local flow details."],"forward_implications":["If the hybrid dual-scale solver is correct, PINN-based microscale flow approximation becomes feasible in periodic fiber geometries, removing the need for mesh-based microscale solvers in tow permeability computation.","NUM and SBM provide a practical route to in-plane permeability for textile composites: under 11 hours of runtime versus 88 to 120 hours for FRM, with consistent $K_{11}$ and $K_{22}$ trends.","The segment-wise micropermeability assignment in SBM does not yield a clear accuracy gain over the uniform NUM assignment in these test cases, so structural variability inside rovings is not the dominant error source for in-plane flow.","Out-of-plane permeability $K_{33}$ remains substantially more variable than in-plane values, with coefficients of variation up to about 99%, so through-thickness flow is the direction where simplified upscaling is least reliable.","The hybrid solver currently needs 20 to 25 minutes of training on a single GPU for a 2D geometry, so scaling to large 3D fibrous models requires distributed domain decomposition and faster optimization."],"supporting_citations":[{"why":"Supplies the PINN formulation and objective that the hybrid solver extends.","marker":"[5]"},{"why":"Provides the hybrid multiscale solver framework this paper builds on.","marker":"[6]"},{"why":"Supplies the periodic boundary conditions and Darcy-law averaging for permeability.","marker":"[7]"},{"why":"Provides the numerical data generation and SVE models for fiber microstructures.","marker":"[12]"},{"why":"Supplies the Stokes-Brinkman solver used for the mesoscale flow in NUM and SBM.","marker":"[14]"},{"why":"Provides the data-driven emulators used for segment-wise micropermeability in SBM.","marker":"[15]"},{"why":"Supplies the Taylor-Hood finite elements used for both reference and mesoscale discretizations.","marker":"[21]"},{"why":"Supplies the Fourier features and gradient-based weight scaling used in PINN training.","marker":"[22]"},{"why":"Provides the gradient-flow stiffness analysis that motivates the weight scaling.","marker":"[26]"}],"fun_headline_variants":["Dual-scale PINN repairs fiber flow blindness in minutes","Hybrid solver weaves scales to predict permeability fast","Scale-bridging network beats stagnant single-scale PINN","Two-scale learning predicts fiber permeability in minutes","Hybrid PINN solver cracks fibrous permeability puzzle"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The hybrid solver's success depends on the mesoscale Stokes-Brinkman solution being a reliable coarse-scale guide and on the hand-picked projection interval $[5\\times10^{-5},\\,5\\times10^{-4}]$ and initial guess $4.5\\times10^{-4}$ bracketing or seeding the true micropermeability; the paper does not test how the result changes if those choices are wrong.","fun_headline_variants_meta":{"raw":{"variants":["Dual-scale PINN repairs fiber flow blindness in minutes","Hybrid solver weaves scales to predict permeability fast","Scale-bridging network beats stagnant single-scale PINN","Two-scale learning predicts fiber permeability in minutes","Hybrid PINN solver cracks fibrous permeability puzzle"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1439,"prompt_tokens":1126,"completion_tokens":313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":742,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":742,"tokens_out":313,"duration_ms":3934,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T20:27:43.131820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the hybrid solver on the 25-fiber geometry with the projection upper bound set below the finite element reference value $2.37\\times10^{-4}$, for instance $K^p_{\\mathrm{UB}}=2.0\\times10^{-4}$; if the predicted permeability lands on the upper bound, then the hand-chosen interval, not the learned flow, is setting the reported accuracy.","supporting_citations":[{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"Supplies the PINN formulation and objective that the hybrid solver extends."},{"cited_title":"A hybrid physics-informed neural network based multiscale solver as a partial differential equation constrained optimization problem","cited_arxiv_id":"2309.04439","evidence_quote":"Provides the hybrid multiscale solver framework this paper builds on."},{"cited_title":"Griebel, M","cited_arxiv_id":null,"evidence_quote":"Supplies the periodic boundary conditions and Darcy-law averaging for permeability."},{"cited_title":"Schmidt, D","cited_arxiv_id":null,"evidence_quote":"Provides the numerical data generation and SVE models for fiber microstructures."},{"cited_title":"Hilden, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Stokes-Brinkman solver used for the mesoscale flow in NUM and SBM."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the data-driven emulators used for segment-wise micropermeability in SBM."},{"cited_title":"Boffi, F","cited_arxiv_id":null,"evidence_quote":"Supplies the Taylor-Hood finite elements used for both reference and mesoscale discretizations."}],"review_version":1}