{"id":"6e7fd1c8-c258-4114-88c0-c4704fabef2b","arxiv_id":"2608.07722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Fourier neural operators reproduce simulated velocity and vorticity fields around immersed-boundary eel swimmers with 3 to 6 percent error, while pressure errors reach 19 percent.","lead":"The paper builds neural-network surrogates that predict the fluid flow fields around swimming eels from one simulation snapshot to the next, in both 2D and 3D. On held-out swimming conditions, the surrogates reproduce velocity and vorticity fields within a few percent, though pressure is less accurate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No persistence baseline or recursive rollout is reported, so the one-step L2 errors do not yet establish FNO predictive skill beyond temporal autocorrelation; the central feasibility claim is under-supported as stated.","rationale":"The paper is a transparent feasibility study with explicit error definitions, honest limitations, and a sensible qualitative finding: velocity and vorticity are predicted much better than pressure. The reader's weakest assumption, that teacher-forced one-step error is meaningful for temporal prediction, is closely related to my concern, but the sharper formulation is the absence of a persistence baseline. Without that baseline, even the one-step numbers do not establish that the FNO has learned anything beyond temporal autocorrelation, which is the minimal condition for the central claim. I do not think this requires rejecting the paper; the qualitative feasibility claim for velocity and vorticity may survive, but the quantitative headline should be read as conditional on the FNO beating a persistence baseline and on the held-out trajectories being used only for final evaluation rather than checkpoint selection. The paper itself recommends stronger validation protocols in Section 7.5, and the lack of a baseline is consistent with that self-identified limitation. The reader's CONDITIONAL verdict remains appropriate, so I recommend no change to the verdict, while adding the persistence test as the specific condition that would settle the concern.","tokens_in":13090,"tokens_out":7296,"duration_ms":75449,"concrete_test":"Compute the persistence error on the same held-out trajectories used for Tables 3 and 5: for every target group, evaluate E_persist = ||Y_{n+1} - Y_n||_L2 / ||Y_{n+1}||_L2 over all held-out temporal pairs and regions (full domain and fluid mask). If the FNO error exceeds the persistence error by less than a meaningful margin, or if the FNO error is not substantially below persistence (e.g., by >1 percentage point for velocity/vorticity), the central feasibility claim is weakened; if the FNO clearly beats persistence, the one-step claim gains support. A second, complementary check is an autoregressive rollout of the FNO over several time steps, comparing accumulated error against the true trajectory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Fourier neural operators can serve as field-resolved temporal surrogates for immersed-boundary swimmer flows, supported by teacher-forced one-step global relative L2 errors of 3.51% (planar) and 3.44%/5.59%/19.2% (volumetric velocity/vorticity/pressure) in Section 6, Tables 3 and 5. Because the input fields at time t_n are fed directly as the prediction input, and the executed time steps are small (planar t: 0.71 to 0.72; volumetric t: 0.97384 to 0.98827), the output fields are strongly correlated with the input fields. A trivial persistence predictor that simply outputs the input fields unchanged would also achieve low relative L2 error on these slowly evolving flows. The paper reports no persistence baseline, no linear-extrapolation baseline, and no recursive rollout, so the reported errors do not yet separate actual learned dynamics from temporal autocorrelation. This matters because the abstract and introduction motivate the surrogate as a replacement for repeated high-fidelity simulation calls, which requires either a demonstrated advantage over persistence and/or stable recursive use; Section 7.4 explicitly limits claims to one-step prediction, but that limitation makes the missing baseline comparison more salient, not less. Without a baseline, the 3.4-3.6% velocity and vorticity errors are not evidence that the neural operator has learned the swimmer flow map rather than the near-identity temporal correlation of the data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops Fourier neural operator (FNO) surrogates for one-step temporal prediction of hydrodynamic fields around immersed-boundary eel swimmers. A planar model jointly predicts two velocity components, scalar vorticity, and pressure, while the volumetric formulation uses three separate target-specific FNOs for velocity, vorticity, and pressure. The models are trained on regular-grid fields exported from adaptive IBAMR simulations, conditioned on swimmer geometry masks and Reynolds number, and evaluated on held-out trajectories. Reported full-domain global relative L2 errors are 3.51% for the planar joint model and 3.44%, 5.58%, and 19.2% for volumetric velocity, vorticity, and pressure, respectively. The paper also reports fluid-region errors, component-wise errors, and physical diagnostics showing that predicted fields violate incompressibility and vorticity-velocity consistency more than the target fields. The authors explicitly frame the work as a feasibility study and list several limitations, including one-step-only evaluation and single-seed training.","tokens_in":13386,"tokens_out":4269,"duration_ms":46245,"significance":"If the evaluation protocol is made sound, this is a useful application-scale benchmark: it brings neural operators to moving-boundary FSI data with multi-channel outputs, parameter conditioning, and genuinely volumetric fields. The paper is honest in reporting pressure weakness, physical inconsistency, and the lack of recursive stability, and its error definitions are clear and reproducible in principle. However, the current evidence does not fully support the central claim of temporal prediction skill, because the headline errors come from a checkpoint-selection protocol that uses the held-out trajectories and because no persistence baseline or rollout experiment is reported. The contribution is best read as a data-generation and baseline benchmark that needs strengthened evaluation before the quantitative claims can be accepted.","major_comments":[{"comment":"The checkpoint-selection protocol uses the held-out trajectories. Section 5.2 states that selection \"minimized the held-out mean per-sample joint full-domain relative L2 loss,\" and Figure 2 shows the monitored held-out curves with the frozen checkpoints marked. The planar split is even labeled \"held-out selection/evaluation\" in Appendix A. Consequently, the errors in Tables 3 and 5 are minima over training epochs on the same set that is later used for reporting, so they are not independent test estimates and are likely optimistic. The authors list the correct protocol in Section 7.5 (\"reserve trajectory-level validation and final test sets\"), but the current paper still presents these numbers as held-out generalization results. This should be fixed by retraining with a separate validation set and reporting final-test errors, or by clearly relabeling the reported numbers as validation-selected estimates and acknowledging the bias quantitatively.","section":"Section 5.2 and Figure 2"},{"comment":"The evaluation is teacher-forced one-step prediction only, and the executed time steps are very small (planar t=0.71 to 0.72; volumetric t=0.97384 to 0.98827). Since the input is the reference field at the previous time step, a trivial persistence predictor that simply copies the input would also achieve low relative L2 error on these slowly evolving flows. The paper reports no persistence baseline, no linear-extrapolation baseline, and no recursive rollout. Therefore the reported 3.44%–3.51% velocity and vorticity errors do not by themselves separate learned dynamics from temporal autocorrelation, which is load-bearing for the claim that the surrogate can \"temporal[ly] predict\" the hydrodynamic fields and eventually replace repeated simulation calls. I request a persistence baseline and at least one short recursive rollout experiment, or a clear statement of why persistence is not a meaningful comparator for this protocol.","section":"Sections 3.2, 6, and 7.4"},{"comment":"All experiments use a single training seed, and Figure 6 explicitly notes that each point is not an uncertainty estimate across seeds. The quantitative claims in Tables 3 and 5 are point estimates from one run. Given that the paper emphasizes reproducible evaluation artifacts, the sensitivity of the reported errors to seed and initialization should be quantified, at least for the smaller planar model or for a subset of volumetric runs. Without this, the reported precision (e.g., 3.51% versus 3.44%) cannot be distinguished from training stochasticity.","section":"Section 7.4 and Table 5"}],"minor_comments":[{"comment":"In the sentence beginning \"Foo all 1000 held-out temporal pairs,\" the word \"Foo\" appears to be a typo and should read \"For.\"","section":"Section 6.1"},{"comment":"The abstract reports the volumetric vorticity error as 5.58% and pressure as 19.2%, while Table 5 lists 5.586% and 19.203%; the rounding is fine, but the abstract's \"5.59%\" in Section 8 should be checked for consistency with Table 5.","section":"Abstract and Table 5"},{"comment":"The planar forward time is reported at batch size 8 while the volumetric times are at batch size 1; the table caption should state this explicitly to avoid an apples-to-oranges comparison.","section":"Table 8"},{"comment":"The repository is announced as \"will be made publicly available soon,\" but the paper's conclusion claims that \"frozen models and reproducible evaluation artifacts\" provide a concrete basis for further work; please clarify the current availability status or provide a permanent snapshot identifier.","section":"Data and Code Availability"},{"comment":"The sentence \"The present masks provide useful geometry conditioning and fluid-region reporting, but a signed-distance field or Lagrangian geometry encoder could better resolve subcell interface location\" is clear, but it would help to state whether the binary occupancy proxy M_e is used only as an input channel or also in the loss weighting.","section":"Section 7.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable application benchmark, and the persistence-baseline and checkpoint-selection issues are fixable within the scope of a revision. The authors are unusually candid about limitations, which is a strength. The reference list includes a 2026 item with a DOI; I did not verify it, but the editors may wish to check that the publication status is accurate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a transparent feasibility study that does something new—field-resolved temporal prediction for immersed-boundary swimmer flows with geometry and Reynolds conditioning—and reports its metrics fairly. The planar and volumetric datasets, the held-out protocol, and the consistency diagnostics are real contributions. The authors do not oversell: they flag that the results are one-step, single-seed, and extrapolated only in Re.\n\nThe main caveat is that the headline errors are teacher-forced one-step predictions on slowly evolving fields. The time steps are tiny (planar 0.71 to 0.72, volumetric ~0.014 s), so a persistence baseline would trivially achieve low relative L2 error. The paper reports no persistence comparison, no linear extrapolation baseline, and no recursive rollout. The physical diagnostics (divergence and vorticity inconsistency) show the model is not simply copying the input—the prediction degrades those differential quantities—but that does not establish it outperforms persistence on the field error. So the 3.5% and 3.4% velocity numbers are plausible but not yet demonstrated to be beyond autocorrelation.\n\nOther soft spots: checkpoint selection used the same held-out trajectories (Section 5.2, Figure 2), so the reported errors are not fully independent test estimates; only one seed is used; pressure error is 19% and the gauge-insensitive diagnostic doesn't fix it; and code/data are announced but not yet released. None of these are fatal for a feasibility study, and the authors explicitly mention several of them in Section 7.4.\n\nWho this is for: anyone building surrogates for moving-boundary FSI or benchmarking neural operators on engineering-scale unsteady flows. It deserves a serious referee; the main requests should be a persistence baseline, at least a short recursive rollout, and checkpoint selection on a validation split separate from the test trajectories. With those additions, the quantitative claims would be much stronger.","headline":"Honest feasibility study of FNOs for swimmer flow fields; the one-step numbers need a persistence baseline before they carry weight.","tokens_in":13916,"tokens_out":2154,"would_cite":true,"duration_ms":20129,"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 Fourier neural operator predicts the full velocity, vorticity, and pressure fields of immersed-boundary eel swimmers one time step ahead, with 3.51% planar and 3.44% volumetric velocity field error.","keywords":["Neural operator","Fourier neural operator","immersed-boundary method","fluid–structure interaction","bio-inspired swimmer","surrogate modeling","computational fluid dynamics","scientific machine learning"],"falsifier":"Run the published frozen checkpoints recursively: start from one held-out state, feed each prediction back as the next input for dozens of steps, and compare the unrolled fields with the simulation; if global relative error grows well above the reported 3.5% / 3.4% levels or the divergence and vorticity-curl residuals amplify within a few steps, the temporal-prediction claim as a practical surrogate would be falsified.","tokens_in":12882,"feed_emoji":"🌊","tokens_out":9379,"duration_ms":79668,"temperature":0.7,"pith_summary":"The paper asks whether a Fourier neural operator (FNO) can replace repeated high-fidelity fluid–structure simulations of a swimming eel by directly predicting the surrounding velocity, vorticity, and pressure fields one time step ahead. Its central claim is yes: a planar FNO jointly predicts all four scalar fields with a 3.51% full-domain global relative $L^2$ error on held-out high-Reynolds-number trajectories, and volumetric models reach 3.44% for velocity, 5.58% for vorticity, and 19.2% for pressure on within-range trajectories. The authors stress that these numbers are for teacher-forced one-step prediction, not recursively unrolled trajectories. If the claim holds, a trained surrogate could make design sweeps, parameter studies, and control loops much cheaper by returning entire distributed fields rather than scalar performance measures.","feed_headline":"Neural operator predicts swimmer flow fields within 3.5%","feed_subtitle":"Field-resolved surrogate for eel-like swimmers works in 2D and 3D; pressure at 19% error is the next target.","key_machinery":"The load-bearing object is the Fourier neural operator (FNO): a network whose layers mix a local pointwise map with a truncated spectral convolution, allowing it to learn maps between functions sampled on regular grids. Each Fourier block executes $v_{\\ell+1} = \\sigma_\\ell\\left(W_\\ell v_\\ell + M_\\ell \\mathcal{F}^{-1}(R_\\ell \\odot \\mathcal{F}(v_\\ell))\\right)$, where $R_\\ell$ is a learned complex multiplier on retained modes. Conditioning on the moving swimmer comes from a binary occupancy mask and a broadcast Reynolds number, with phase features $\\sin\\phi, \\cos\\phi$ added in the volumetric case, and the temporal task is the reference-state map from fields at $t_n$ to fields at $t_{n+1}$. The volumetric formulation is three target-specific FNOs sharing one multichannel input rather than a joint seven-output network.","core_discovery":"On its own terms, the paper demonstrates that a Fourier neural operator can act as a field-resolved surrogate for the hydrodynamic fields of an immersed-boundary swimmer. A planar joint model maps the current velocity, vorticity, pressure, geometry mask, Reynolds number, and time step to the next velocity, vorticity, and pressure fields, and three volumetric models each predict one target group from a common 15-channel input. Across five held-out planar trajectories at Reynolds numbers 7300–8500, the full-domain global relative $L^2$ error of the joint prediction is 3.51%. Across five held-out volumetric trajectories at Reynolds numbers 750–6200, the errors are 3.44% for velocity, 5.58% for vorticity, and 19.2% for pressure. Pressure is the least accurate field, and the planar model's predicted divergence and vorticity-curl inconsistency are larger than the reference fields', so the authors identify physical consistency and pressure accuracy as the main development priorities.","pith_inferences":["If the one-step map is stable under feedback, an obvious extension is to train the same architecture with a rollout loss or a recurrent head so it can be unrolled over many time steps; the paper neither claims nor tests that.","Because removing the spatial mean barely changes the pressure error, the 19% volumetric pressure error is likely spatial structure rather than a gauge offset, so force-relevant integrals such as thrust or pressure drag might be far more accurate than the pointwise norm suggests.","A cheap testable modification is to add an incompressibility penalty or a curl–vorticity consistency term to the training loss; the paper's own diagnostics show the surrogate's divergence and vorticity-curl residuals are the main physics violations.","The conditioning recipe (geometry mask plus Reynolds number) could be pushed toward unseen gaits and body shapes with a latent geometry encoder, since the current study varies only Reynolds number for one eel geometry."],"forward_implications":["A single trained surrogate returns distributed velocity, vorticity, and pressure fields for a new Reynolds number at the cost of one forward pass, which is the ingredient a design-optimization or control loop would need to replace many repeated FSI solves.","The volumetric models generalize across the tested Reynolds range (750–6200) and the planar model extrapolates mildly to 7300–8500, with error growing gradually rather than collapsing.","Velocity and vorticity are accurate enough for wake-structure analysis, while the 19.2% volumetric pressure error means pressure-derived quantities such as loads cannot be taken from the surrogate without dedicated treatment.","The reported accuracy is the accuracy of a one-step transition from a known state; it does not by itself certify multi-step temporal prediction.","Global field accuracy does not imply differential consistency, since the planar surrogate's divergence and vorticity-curl residuals exceed those of the reference fields."],"supporting_citations":[{"why":"introduces the immersed-boundary method whose Eulerian–Lagrangian coupling generates the simulated fields.","marker":"[1]"},{"why":"provides the adaptive second-order immersed-boundary discretization used for the high-fidelity data.","marker":"[2]"},{"why":"supplies the software infrastructure that produced the trajectory datasets.","marker":"[4]"},{"why":"introduces the Fourier neural operator architecture on which the surrogates are built.","marker":"[5]"},{"why":"provides the unified constraint-based fluid–structure interaction formulation used for deforming swimmer bodies.","marker":"[15]"}],"fun_headline_variants":["Neural operator predicts swimmer flows at 3.5% error","AI surrogate for swimmer flows: 3.5% error, pressure lags","Field-resolved neural net: 3.5% error, 19% for pressure","Fourier neural operator learns 2D/3D swimmer hydrodynamics","Swimmer flow AI: velocity under 5%, pressure next target"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that teacher-forced one-step evaluation, where the true fields at time $t_n$ are given as input and the model predicts $t_{n+1}$, is a meaningful measure of temporal prediction; the paper does not demonstrate stable recursive rollout.","fun_headline_variants_meta":{"raw":{"variants":["Neural operator predicts swimmer flows at 3.5% error","AI surrogate for swimmer flows: 3.5% error, pressure lags","Field-resolved neural net: 3.5% error, 19% for pressure","Fourier neural operator learns 2D/3D swimmer hydrodynamics","Swimmer flow AI: velocity under 5%, pressure next target"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000286,"raw_usage":{"total_tokens":1692,"prompt_tokens":965,"completion_tokens":727,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":623}},"tokens_in":581,"tokens_out":727,"duration_ms":6762,"temperature":1.0,"reasoning_tokens":623,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:21:27.126179+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published frozen checkpoints recursively: start from one held-out state, feed each prediction back as the next input for dozens of steps, and compare the unrolled fields with the simulation; if global relative error grows well above the reported 3.5% / 3.4% levels or the divergence and vorticity-curl residuals amplify within a few steps, the temporal-prediction claim as a practical surrogate would be falsified.","supporting_citations":[{"cited_title":"Griffith, Richard D","cited_arxiv_id":null,"evidence_quote":"provides the adaptive second-order immersed-boundary discretization used for the high-fidelity data."},{"cited_title":"IBAMR: Immersed boundary adaptive mesh refinement software infras- tructure.https://ibamr.github.io/, n.d","cited_arxiv_id":null,"evidence_quote":"supplies the software infrastructure that produced the trajectory datasets."},{"cited_title":"Griffith, and Neelesh A","cited_arxiv_id":null,"evidence_quote":"provides the unified constraint-based fluid–structure interaction formulation used for deforming swimmer bodies."}],"review_version":1}