{"id":"5ce97c1e-a1df-44b3-81dc-d7af87e240dc","arxiv_id":"2607.04491","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Solver-agnostic PINN training yields four RANS closures that generalize across six bluff-body shapes at Re=10^4, with a Reynolds-force model reaching ~8.5% LOSO drag error.","lead":"A PINN trains RANS turbulence closures without embedding a CFD solver, then freezes them for use in a standard finite-element code across six bluff-body wakes. The approach screens models in minutes and, under leave-one-shape-out tests, beats a steady SST k–ω baseline on mean flow and drag.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"LOSO outperformance holds in the reported regime; the load-bearing soft spot is the untested premise that single-Re symmetric-mean geometry transfer extends to other Re or full-domain asymmetric flows.","rationale":"The reader’s strongest claim accurately restates what the tables and figures establish inside the six-shape, single-Re, half-domain setting, and the reader’s weakest assumption correctly isolates the untested transfer premise that underwrites the broader “generalizable across shapes” language. No stronger internal inconsistency (e.g., in the PDE-consistency argument of §4.1, the stabilization ablation of §4.2, or the energy-neutral projection of Eq. 3.12) undermines the reported LOSO numbers themselves. The concrete multi-Re or full-domain check is exactly the experiment the paper’s own limitations section flags as next; until it is run the CONDITIONAL verdict and medium correctness risk remain appropriate. No change to the reader’s assessment is warranted.","tokens_in":26024,"tokens_out":573,"duration_ms":24972,"concrete_test":"Retrain all four closures under the identical LOSO protocol at a second Re (e.g. 5e3 or 5e4) on the same six shapes, or drop the half-domain symmetry BC for at least the circle and square at Re=1e4 and redeploy full-domain; if M4 LOSO mean relative drag error exceeds ~15% or mean-velocity L2 errors roughly double those in Table 3, the transfer premise fails and the generalizability claim must be narrowed to the demonstrated regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim (strict LOSO outperformance of four PINN-trained closures vs SST on mean fields/drag at Re=10^4) is supported by Tables 3–4 and the field comparisons. The load-bearing vulnerability is the paper’s framing of these results as evidence of generalizable closures across shapes (§1, title, abstract, §4.5–5). That framing rests on the untested premise that geometry transfer demonstrated only for steady, spanwise-averaged, half-domain-symmetric 2-D means at one Re will hold at other Re or without the symmetry BC. The paper itself notes the single-Re and half-domain limits and invokes analogy to prior cylinder Re-transfer work rather than new multi-Re or full-domain experiments. If the learned maps (especially the shared C and length-scale nets) are tuned to the fixed-Re balance or to the forced symmetry used in both PINN training and FEM deployment, the claimed cross-shape generality is regime-specific.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes training RANS turbulence closures inside a physics-informed neural network so that the inverse problem is mesh-free and solver-agnostic: the RANS residual is imposed by automatic differentiation, with no CFD solver, mesh, or adjoint in the training loop. Four closures are developed on equal footing—three realizable tensor-basis Reynolds-stress models (local M1; non-local transported-k with algebraic length scale M2; non-local with a learned k-free length scale M3) and a structure-preserving Reynolds-force model M4 that targets F=−∇·τ with an energy-neutral residual projection. All four are trained on six 2-D bluff-body wakes at Re=10^4 (DNS, with a PIV demonstration) and deployed frozen in an independent finite-element RANS solver, stabilized by input-gradient smoothing and a Lipschitz penalty. Under a strict leave-one-shape-out protocol, all four substantially outperform a steady SST k–ω baseline; M3 is most accurate on the stress fields, while M4 generalizes best on mean velocity and drag (mean LOSO relative drag error ~8.5%).","tokens_in":26440,"tokens_out":1391,"duration_ms":22150,"significance":"If the reported LOSO gains and stable a-posteriori FEM deployment hold, the work is a meaningful methodological contribution to data-driven RANS: it shows that PDE-consistent, solver-agnostic PINN training can produce transferable stress and force closures that clear the usual a-priori/a-posteriori gap without embedding a solver in the loop, at training costs of minutes per hypothesis on one GPU. Strengths include the strict LOSO design across six distinct wakes, the a-priori vs a-posteriori contrast (Fig. 4), the stabilizer ablation (Fig. 5), the PIV-trained demonstration (Fig. 6), the FIML/SA comparison (Fig. 9), and quantitative field and drag tables (Tables 3–4). The dual stress/force treatment within one trainer is also useful. The result is of clear interest to the turbulence-modelling and scientific-ML communities, even though the tested regime is steady, spanwise-averaged, symmetric-mean wakes at a single Re.","major_comments":[{"comment":"The title, abstract, and §1/§4.5–5 frame the result as geometry-generalizable closures, but all training and LOSO tests are at fixed Re=10^4 on steady, spanwise-averaged means with half-domain symmetry in FEM deployment (§2, §3.3). §5 correctly lists single-Re and half-domain limits and invokes analogy to prior cylinder Re-transfer work rather than new multi-Re or full-domain tests. The central LOSO claim at this Re is supported by Tables 3–4; the broader framing should be tightened so that “generalizable across shapes” is explicitly scoped to the tested regime, and any Re-transfer expectation is stated as a hypothesis, not an implied result.","section":null},{"comment":"§3.3 and Fig. 5 establish that input-gradient Helmholtz smoothing (δ_f=c h, c=1) is decisive for Picard convergence of the frozen closure. Because this filter is applied to velocity gradients before they enter C, the deployed operator is not identical to the raw network map learned in the PINN. The paper should quantify how much a-posteriori accuracy (Tables 3–4, field figures) depends on this filter width—e.g., a short sensitivity on c or an explicit statement that reported errors include this deployment regularizer—so that the claimed transfer of the learned closure is not conflated with mesh-scale smoothing.","section":null},{"comment":"Main quantitative claims rest on SST k–ω as the sole baseline in Tables 3–4 and the six-way field figures; the FIML+SA comparison (Fig. 9) is only in-sample on the cylinder. For a JFM-level claim of substantial outperformance of data-driven closures, at least one additional classical nonlinear/eddy-viscosity or algebraic stress baseline (or a brief note why SST alone is the appropriate industrial reference) would strengthen the interpretation that gains come from the PINN-trained forms rather than from beating a single linear Boussinesq model on massively separated wakes.","section":null}],"minor_comments":[{"comment":"§3.1, Eq. (3.12): the energy-neutral projection R_⊥ is central to M4’s structure preservation; a short remark on whether the projection is applied only at training, only at deployment, or both would help reproducibility.","section":null},{"comment":"Table 3 reports mean relative-L2 over six shapes but not per-shape standard deviation or range; adding a brief dispersion measure (or pointing more explicitly to the per-shape field figures) would clarify whether LOSO averages are driven by one hard geometry (e.g., the ellipse in Fig. 12).","section":null},{"comment":"§3.2 training cost comparison to the authors’ prior OpenFOAM adjoint (~1 day) is useful; stating mesh size / core count for that baseline next to the PINN GPU timing would make the “orders of magnitude” claim fully checkable.","section":null},{"comment":"Notation: q7/q8 and the bounded limiter (Table 1) are clear, but C′ for the length-scale net is easy to miss in the text; a single sentence in §3.1 listing which methods use C′ would help.","section":null},{"comment":"Appendix B Table 5 (training-recipe ablation) is valuable; a one-line pointer from §4.1 would make it easier to find.","section":null},{"comment":"Minor typography: “Focus on Fluids articles must not exceed this page length” and similar production banners remain in the manuscript text and should be removed before typesetting.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Solid, carefully executed JFM-level contribution on methodology and LOSO evidence. I would not block on the single-Re/symmetry scope if the authors tighten the framing as requested; the technical core (PDE-consistent PINN trainer, stable FEM deployment, four closures, LOSO tables) is sound. Fit for JFM is appropriate given the turbulence-closure focus and quantitative a-posteriori tests."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful result is practical: they train four structured RANS closures (three realizable tensor-basis stress models plus a structure-preserving force model) inside a PINN with no CFD solver in the inverse loop, freeze them, and get stable FEM deployments that beat steady SST k–ω under strict leave-one-shape-out across six bluff-body wakes at Re=10^4. Force model M4 hits ~8.5% mean relative LOSO drag error; learned-ℓ M3 is best on the stresses. That combination is the novelty—not tensor bases (Ling), not Reynolds force (Cruz), not FIML, not single-case PINN–RANS, not their own prior cylinder Re-transfer work—but multi-geometry, solver-agnostic training plus a-posteriori FEM and LOSO numbers.\n\nWhat they do well is the chain of evidence. PDE residual in training is shown to matter for transfer (a-priori fit collapses on deployment; supervised PINN does not). Stabilizer ablation isolates input-gradient smoothing as the decisive piece. PIV-patched training works for the in-plane models. Tables 3–4 and the field figures are quantitative and honest about which model wins which field. Cost claim is concrete: minutes on one GPU versus ~day-scale adjoint. Citations are appropriate; free parameters (λ_lip, length-scale bounds, filter width) are regularizers, not the reported drag/velocity wins.\n\nSoft spots are real but stated by the authors and proportional. Single Re, spanwise-averaged 2-D means, half-domain symmetry: the paper frames “generalizable across shapes” and expects Re transfer by analogy to prior cylinder work, without multi-Re or full-domain tests here. Absolute stress errors remain modest; ellipse is the weak geometry. No public code. None of that voids the LOSO outperformance inside the reported regime; it bounds the claim.\n\nThis is for people who build or deploy data-driven RANS closures and care about inverse cost, architecture freedom, and a-posteriori stability. It deserves a serious referee. I would engage with it and expect revision on scope language and artifacts, not rejection of the core result.","headline":"Solid methods paper: PINN trains multi-geometry RANS closures without a solver in the loop, and LOSO FEM results beat SST at Re=10^4; the generality claim is regime-bound but the evidence inside that regime is real.","tokens_in":27066,"tokens_out":567,"would_cite":true,"duration_ms":6052,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Turbulence closures trained without a CFD solver in the loop generalize across bluff-body shapes and beat SST k–ω on drag and mean flow.","keywords":["turbulence closure","physics-informed neural networks","RANS","bluff-body wakes","Reynolds force","leave-one-shape-out","solver-agnostic training"],"falsifier":"Train the same four closures at Re = 10^4, freeze them, and deploy them without retraining on a bluff body at a substantially different Reynolds number (e.g., 5 × 10^3 or 10^5) or on a full-domain asymmetric mean flow with vortex shedding or lift; if mean-velocity or drag errors rise to or above the SST baseline, the claimed transfer fails.","tokens_in":26905,"feed_emoji":"🌊","tokens_out":1049,"duration_ms":8128,"temperature":0.7,"pith_summary":"Data-driven turbulence models are usually fit by repeatedly running a CFD solver inside the optimization, which ties the model to one mesh and solver and forces every training iterate to converge. This paper instead trains the closure inside a physics-informed neural network: the Reynolds-averaged Navier–Stokes residual is enforced by automatic differentiation, so training is mesh-free, fully differentiable, and independent of any external solver. Only the final frozen closure must be stable when plugged into a standard finite-element code. Four closures are developed—three that reconstruct the Reynolds-stress tensor on a realizable tensor basis (local, non-local with transported kinetic energy, and the same with a learned length scale) and one that models the Reynolds force directly. Trained on six distinct two-dimensional bluff-body wakes at Re = 10^4 and tested under a strict leave-one-shape-out protocol, all four beat a steady SST k–ω baseline; the force model recovers drag to about 8.5% mean relative error while the learned-length-scale model is most accurate on the stresses. The same trainer also works from patched experimental PIV data, opening geometries that DNS cannot reach.","feed_headline":"PINN-trained closures beat SST k–ω across unseen bluff shapes","feed_subtitle":"No CFD solver in the training loop; force model recovers drag to ~8.5% on held-out wakes","key_machinery":"Solver-agnostic PINN training: the shared closure network is optimized jointly with per-shape mean-field networks by soft residual penalties (momentum, continuity, and, when used, the k-transport equation) obtained by automatic differentiation; no forward RANS solve occurs during training, so intermediate iterates need not be solver-stable and no closure-specific adjoint is required.","core_discovery":"A PINN that imposes the RANS residual by automatic differentiation can train transferable turbulence closures without any CFD solver, mesh, or adjoint in the loop; the resulting frozen models, stabilized by input-gradient smoothing and a Lipschitz constraint, deploy stably in a standard finite-element solver and, under leave-one-shape-out testing across six bluff-body wakes, substantially outperform SST k–ω on mean velocity and drag (force model ~8.5% LOSO drag error) while the learned-length-scale stress model leads on the stress fields.","pith_inferences":["The same residual-based trainer could be applied to three-dimensional or unsteady RANS/URANS closures once full-domain data become available.","Because the force model never needs a realizability cap, it may be the more scalable route when the stress tensor itself is not required for design.","If the single-Re geometry transfer holds, multi-shape training sets could become a practical alternative to classical calibration on a handful of canonical flows."],"forward_implications":["New closure hypotheses can be screened in minutes on one GPU instead of days of adjoint or ensemble solves.","A single force-based model can recover mean velocity and drag across unseen bluff shapes more accurately than classical two-equation RANS.","Patched experimental PIV can replace DNS for training, extending data-driven closures to geometries that cannot be simulated at high fidelity.","Stable a-posteriori deployment of neural closures becomes routine once input-gradient smoothing and a Lipschitz penalty are applied."],"fun_headline_variants":["PINN trains solver-free turbulence closures that beat SST on unseen bluffs","Mesh-free PINN RANS closures generalize across bluff shapes leave-one-out","Force PINN closure hits ~8.5% drag error on held-out bluff wakes","Learned-length PINN stress models top SST on Reynolds fields across shapes","Solver-agnostic PINN yields transferable RANS closures for bluff wakes"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The claim that geometry generalization shown at one Reynolds number on steady, spanwise-averaged, half-domain-symmetric wakes will also hold at other Reynolds numbers and for flows whose mean is not symmetric.","fun_headline_variants_meta":{"raw":{"variants":["PINN trains solver-free turbulence closures that beat SST on unseen bluffs","Mesh-free PINN RANS closures generalize across bluff shapes leave-one-out","Force PINN closure hits ~8.5% drag error on held-out bluff wakes","Learned-length PINN stress models top SST on Reynolds fields across shapes","Solver-agnostic PINN yields transferable RANS closures for bluff wakes"]},"model":"grok-4.5","effort":"low","cost_usd":0.003736,"raw_usage":{"total_tokens":1300,"prompt_tokens":920,"num_sources_used":0,"completion_tokens":91,"cost_in_usd_ticks":37360000,"prompt_tokens_details":{"text_tokens":920,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":289,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":920,"tokens_out":91,"duration_ms":3538,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T06:58:22.535176+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the same four closures at Re = 10^4, freeze them, and deploy them without retraining on a bluff body at a substantially different Reynolds number (e.g., 5 × 10^3 or 10^5) or on a full-domain asymmetric mean flow with vortex shedding or lift; if mean-velocity or drag errors rise to or above the SST baseline, the claimed transfer fails.","supporting_citations":[],"review_version":2}