{"id":"83f8ccfc-8a48-41db-b697-6724f24a3b72","arxiv_id":"2607.11576","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PINNs that embed incompressible Navier-Stokes residuals with sparse experimental velocities reconstruct high-fidelity hemodynamic fields and indicators more accurately than pure CFD or data-only networks on FDA nozzle and aneurysm cases.","lead":"Physics-informed neural networks fuse sparse velocity measurements with the Navier-Stokes equations to reconstruct high-resolution flow, pressure, and wall shear stress. This can raise the clinical value of under-resolved 4D flow MRI and PIV for vascular disease assessment.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"The central claim of superiority over CFD and pure data-driven methods rests on a quasi-steady residual that omits ∂u/∂t for pulsatile aneurysm flow, yet no quantitative check of that omission is supplied.","rationale":"The Reader correctly isolates the quasi-steady residual as the weakest assumption that underpins the central claim. The FDA nozzle is genuinely steady, so the claim holds there; the aneurysm is the only place where the claim is stress-tested against pulsatile data, and there the physics residual is deliberately incomplete. Because the paper already flags the issue in §4.1 and still asserts the superiority statement for both cases, the concern is load-bearing rather than peripheral. The concrete residual-magnitude test is inexpensive (it re-uses existing FEM or multi-frame 4D-flow data) and would either confirm that the omitted term is negligible or force a revision of the claim. No stronger internal inconsistency appears; the hand-tuned weights and missing code are secondary engineering limitations already noted by the Reader. Hence the verdict remains CONDITIONAL, with the same high-confidence medium-risk profile.","tokens_in":24649,"tokens_out":808,"duration_ms":8159,"concrete_test":"At one of the reported aneurysm snapshots (e.g. t=0.34 s), evaluate the pointwise magnitude of the neglected term |∂u/∂t| from the high-resolution FEM solution (or from finite differences of successive 4D-flow frames) and compare it with the retained terms |(u·∇)u| and |(1/Re)Δu| on the same collocation set used for L_PDE. If the median ratio |∂u/∂t| / |(u·∇)u| exceeds ~0.2, retrain an otherwise identical PINN that includes the unsteady residual and recompute the velocity/pressure/WSS errors of Table 6; a degradation larger than the reported CFD-to-PINN gap would falsify the superiority claim under the quasi-steady residual.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (abstract and §4) is that the PINN synergy yields results closer to ground truth than standard CFD or pure data-driven approaches. For the aneurysm, the residual used in L_PDE is the steady form of the Navier–Stokes equations (Eqs. 1a–b and 2a–b) evaluated at a single fixed time t (explicitly stated in §3.2.1: “by fixing the time … to t=2.024 s” and again for the 4D-flow cases at isolated snapshots). The unsteady term ∂u/∂t is therefore never present in the physics loss, even though the inflow is pulsatile (Fig. 13b) and the Reynolds number is ~1500. The paper itself lists this quasi-steady assumption as a limitation (§4.1), yet the claim of superiority is still asserted for the aneurysm results (Figs. 20–24, Table 6). If the neglected acceleration term is comparable in magnitude to the retained convective or viscous terms at the chosen instants, the physics residual is systematically incorrect; any apparent improvement over CFD or pure data-driven baselines could then be an artifact of that modeling error rather than genuine synergy. No residual-magnitude comparison or full-unsteady PINN control is reported, so the load-bearing condition for the claim remains unverified on the more complex of the two benchmarks.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes a PINN framework that embeds the incompressible Navier–Stokes residuals (Eqs. 1–3, 6–7) together with sparse velocity measurements into a composite loss (Eq. 5) to reconstruct high-resolution velocity, pressure and wall-shear-stress fields from under-resolved experimental data. Two benchmarks are examined: the laminar FDA nozzle (Re = 500), trained and tested against both FEM solutions and PIV measurements (FDA-CFD 1–7, FDA-EXP 1–3), and a patient-derived aneurysm model (Re ≈ 1500) trained on FEM snapshots and on in-vitro 4D-flow MRI (AN-CFD, AN-4DMRI weight sweeps). Systematic ablations of loss weights, MSLE terms, unit-vector losses and collocation density are reported, together with quantitative error metrics (Ez, Eu, kinetic energy, enstrophy) and visual comparisons of velocity, pressure, vorticity and WSS. The central claim is that the data–physics synergy yields reconstructions closer to ground truth than pure CFD or pure data-driven baselines.","tokens_in":25067,"tokens_out":1193,"duration_ms":11820,"significance":"If the claimed superiority holds, the work supplies a practical, mesh-free post-processing pipeline that can recover pressure and wall shear stress from sparse 4D-flow MRI or PIV without full CFD personalization. The multi-configuration ablations (Tables 2–5), explicit residual definitions, and dual validation against both FEM and experimental data constitute a reproducible demonstration of PINN utility for hemodynamic indicators. The FDA-nozzle results in particular are carefully controlled and already of interest to the validation community. The aneurysm application, while more ambitious, remains limited by the quasi-steady residual assumption that the authors themselves flag.","major_comments":[{"comment":"§3.2.1 and §3.2.2 (and Limitations §4.1): for the aneurysm the physics residual L_PDE is the steady form of the Navier–Stokes equations (Eqs. 1a–b, 2a–b) evaluated at isolated fixed times (t = 2.024 s for FEM; discrete snapshots for 4D-flow). The unsteady term ∂u/∂t is therefore omitted even though the inflow is pulsatile (Fig. 13b) and Re ≈ 1500. No residual-magnitude comparison or full-unsteady PINN control is supplied. Because the abstract and §4 assert superiority over CFD and pure data-driven methods for this more complex case, the quasi-steady assumption is load-bearing; either a quantitative check of the neglected acceleration term or a clear restriction of the superiority claim to the steady FDA nozzle is required.","section":"§3.2.1, §3.2.2, §4.1"},{"comment":"Abstract and §4 claim that PINN results “align more closely with ground truth … than standard CFD or pure data-driven approaches.” For the aneurysm the only pure-data control is AN-4DMRI-4 (Table 5, PDE weights set to zero). No corresponding pure-data ablation is reported for the FDA-EXP series, and the CFD baselines are themselves subject to modeling choices (VMS-LES, mesh, boundary conditions). A single, consistently defined pure-data baseline for both benchmarks would make the superiority statement falsifiable.","section":"Abstract, §4, Table 5"}],"minor_comments":[{"comment":"Table 1 lists three different characteristic diameters and viscosities; a short sentence clarifying which length scale is used for the aneurysm Re would avoid ambiguity.","section":"Table 1"},{"comment":"Eq. (11) defines WSS; the subsequent wall-averaged plots (Figs. 8d, 11c, 23) would be clearer if the averaging surface and the precise post-processing library (nisaba vs. Paraview) were stated once in the methods.","section":"Eq. (11), Figs. 8d, 11c, 23"},{"comment":"Several figures (e.g., Fig. 7 residual maps, Fig. 22 vorticity) use log-scale color bars without explicit units or reference values; adding a common color-bar range across panels would aid comparison.","section":"Figs. 7, 22"},{"comment":"The MSLE terms in Eqs. (6) and (9) are introduced without a reference or a short derivation of the +1 offset; a one-sentence justification would help readers unfamiliar with the device.","section":"Eqs. (6), (9)"},{"comment":"Typographical inconsistencies appear in the graphical abstract (“Weusephysics-informed…”) and in a few figure captions; a final proof-reading pass is needed.","section":"Graphical abstract, figure captions"}],"recommendation":"major_revision","confidential_remarks":"The quasi-steady residual is the single most important technical risk. If the authors can either (i) quantify that |∂u/∂t| is negligible at the chosen instants or (ii) rephrase the aneurysm claims as exploratory, the paper becomes a solid contribution suitable for a numerical-methods or biomechanics journal. Without that clarification the superiority claim for the pulsatile case remains unsupported."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The paper is a careful, multi-ablation engineering job that shows PINNs can recover velocity, pressure and WSS from sparse PIV and 4D-flow-MRI better than pure CFD or pure data on two standard hemodynamics cases. That is the useful takeaway.\n\nWhat is actually new is not the idea of PINNs for blood flow (Kissas and Sarabian already did that) but the concrete pipeline: MSLE terms for low-magnitude residuals, unit-vector loss, zero-mean pressure penalty, multi-section sampling, and systematic weight sweeps (FDA-CFD 1–7, FDA-EXP 1–3, AN-CFD/AN-4DMRI). On the FDA nozzle the comparisons are clean; PINNs trained on PIV beat the FEM baseline against the experimental data, and the pressure and wall-averaged WSS look sensible. The aneurysm results recover plausible vorticity paths that sit closer to the higher-resolution 4D-flow set than the training data alone, and they give non-invasive pressure without ever seeing pressure labels. The error tables (Ez, Eu, kinetic energy, enstrophy) and the visual side-by-sides are honest.\n\nThe soft spots are real but already listed by the authors. The quasi-steady residual (no ∂u/∂t, time fixed) is used for the pulsatile aneurysm; that is a modeling approximation whose magnitude is never checked against the retained terms. It weakens the strongest claim of “synergy beats CFD and pure data” on the more complex case, yet it does not invalidate the FDA results or the fact that physics regularization still improves the sparse-data reconstructions. Hyper-parameters are hand-tuned, WSS remains dispersed near the inlet, only laminar flow is treated, and no code is released. None of these are hidden; they simply keep the paper from being unconditional.\n\nMath and citations look solid; the Navier–Stokes residuals are standard and the circularity burden is low. This is for people who already work on 4D-flow post-processing or PINN hemodynamics and want a reproducible-enough recipe with clear ablations. I would send it to referees; they will ask for an unsteady residual check or automated weight selection, but the core evidence is already there.","headline":"Solid engineering of a PINN pipeline for sparse flow data; the quasi-steady residual is a real but already-flagged soft spot that does not erase the FDA gains or the transparent ablations.","tokens_in":25667,"tokens_out":562,"would_cite":true,"duration_ms":6885,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M32","76Z05","68T07"],"pacs":[],"model":"grok-4.5","headline":"Physics-informed neural networks turn sparse, noisy blood-flow velocity measurements into high-resolution velocity, pressure and wall-shear-stress fields that match ground truth better than pure CFD or pure data fitting.","keywords":["PINNs","4D flow MRI","Computational hemodynamics","Scientific machine learning","Wall shear stress","Navier-Stokes reconstruction"],"falsifier":"Train the same PINN architecture on a fully time-resolved, high-resolution ground-truth CFD solution of the aneurysm (including the unsteady term) and check whether the reconstructed wall-shear-stress and enstrophy time series still match the ground truth within the error levels reported for the quasi-steady runs.","tokens_in":25556,"feed_emoji":"🩸","tokens_out":864,"duration_ms":9456,"temperature":0.7,"pith_summary":"Clinical imaging methods such as 4D flow MRI and particle-image velocimetry give incomplete or noisy pictures of blood velocity, especially near vessel walls. Those gaps make derived quantities such as wall shear stress unreliable for diagnosis. This paper shows that a neural network trained simultaneously on the measured velocities and on the residual of the incompressible Navier–Stokes equations can fill in the missing information. On both a standard nozzle benchmark and a patient-derived aneurysm model the physics-regularized reconstructions recover pressure (never measured) and produce smoother, more accurate wall-shear-stress maps than either a pure CFD simulation or a network trained only on the data. The result is a mesh-free post-processing tool that upgrades under-resolved clinical flow measurements into spatially resolved hemodynamic indicators.","feed_headline":"PINNs upgrade sparse blood-flow scans into full pressure and stress maps","feed_subtitle":"Physics residuals fill gaps left by 4D MRI and PIV, beating pure CFD or data-only fits","key_machinery":"The composite PINN loss: mean-squared and mean-squared-logarithmic residuals of mass and momentum conservation, boundary-condition residuals, plus data fidelity terms on velocity magnitude and direction; pressure is recovered a posteriori from the momentum residual without ever being supplied as training data.","core_discovery":"Embedding the steady incompressible Navier–Stokes residuals directly into the training loss of a neural network allows sparse experimental velocity samples (PIV or 4D flow MRI) to be reconstructed into high-resolution velocity, pressure and wall-shear-stress fields that agree more closely with ground-truth observations than either classical CFD or pure data-driven fitting.","pith_inferences":["If the quasi-steady approximation holds for moderate pulsatility, the method could be applied frame-by-frame to full cardiac-cycle 4D flow MRI without enlarging the network architecture.","Automatic balancing of the multi-term loss (rather than manual weight selection) would be the next practical step before routine clinical use.","Extending the residual to include a simple turbulence model or the unsteady term would test whether the same framework remains competitive in transitional or highly unsteady regimes."],"forward_implications":["Sparse 4D flow MRI acquisitions can be upgraded to clinically usable wall-shear-stress and pressure maps without additional imaging time.","Pressure, never measured by phase-contrast MRI, becomes available as a free by-product of the physics residual.","Mesh generation and geometry segmentation steps required by classical CFD can be skipped for post-processing of clinical velocity data.","The same loss construction can be reused on other under-resolved experimental modalities (PIV, Doppler ultrasound) once fluid properties and approximate boundary locations are known."],"fun_headline_variants":["PINNs fuse Navier-Stokes with sparse PIV and 4D MRI for high-res flow maps","Physics residuals lift under-resolved blood-flow scans to full pressure fields","Neural nets plus incompressible NS reconstruct wall shear from sparse velocity data","PINN recovers pressure and WSS from noisy experimental flow measurements","Steady NS residuals turn sparse scans into resolved hemodynamic indicators"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Treating the Navier–Stokes equations as quasi-steady (no time derivative) still yields accurate reconstructions for pulsatile aneurysm flow when only a few discrete time snapshots are given to the network.","fun_headline_variants_meta":{"raw":{"variants":["PINNs fuse Navier-Stokes with sparse PIV and 4D MRI for high-res flow maps","Physics residuals lift under-resolved blood-flow scans to full pressure fields","Neural nets plus incompressible NS reconstruct wall shear from sparse velocity data","PINN recovers pressure and WSS from noisy experimental flow measurements","Steady NS residuals turn sparse scans into resolved hemodynamic indicators"]},"model":"grok-4.5","effort":"low","cost_usd":0.004622,"raw_usage":{"total_tokens":1393,"prompt_tokens":838,"num_sources_used":0,"completion_tokens":99,"cost_in_usd_ticks":46220000,"prompt_tokens_details":{"text_tokens":838,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":456,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":838,"tokens_out":99,"duration_ms":4929,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T04:39:09.774631+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the same PINN architecture on a fully time-resolved, high-resolution ground-truth CFD solution of the aneurysm (including the unsteady term) and check whether the reconstructed wall-shear-stress and enstrophy time series still match the ground truth within the error levels reported for the quasi-steady runs.","supporting_citations":[],"review_version":1}