{"id":"f595acf2-ebd3-48a4-93b4-33a42a8602f9","arxiv_id":"2501.17987","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A SIREN neural network with hand-tuned frequency scaling reconstructs pressure from noisy PIV/LPT data with lower error than OS-MODI and Green's function integration, and needs no mesh.","lead":"This paper reconstructs pressure fields from velocimetry data with a SIREN neural network that matches the measured pressure gradient without any mesh. On turbulence benchmarks the network matches or beats two established integration methods and filters noise by restricting the network's frequency band, though real-data validation is limited to one qualitative experiment.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Architectural noise filtering presumes frequency-separable noise; realistic PIV noise is correlated and boundary-dominated, so the central robustness claim is not yet established.","rationale":"The reader's weakest_assumption correctly identifies spectral separability of noise as load-bearing: the noise-filtering advantage claimed in the abstract and Section 3 is only demonstrated with synthetic Gaussian noise on a single dataset. My stress-test converges on the same point and sharpens it by emphasizing the spatial correlation and boundary concentration of actual PIV measurement errors. The paper is internally consistent and transparent about many limitations — the experimental validation is qualitative, the extrapolated solution is flagged as unreliable, and the c in [2,3] rule is acknowledged as empirical. Those caveats do not contradict the central claim for the tested synthetic regime, but they leave the generalization to realistic noisy environments unverified. This does not warrant rejection; the CONDITIONAL verdict stands because the evidence is promising but incomplete. I agree with the reader that code release and comparisons with error bars, the RBF baseline, and realistic noise would move the paper closer to acceptance; the missing realistic-noise test is the most decisive single check because the noise-filtering property is the paper's distinctive novelty beyond mesh-freeness.","tokens_in":18259,"tokens_out":4501,"duration_ms":49326,"concrete_test":"Generate a realistic PIV measurement from the JHTDB isotropic turbulence velocity field used in §3.1: render synthetic particle images, perform cross-correlation PIV with 16×16-pixel interrogation windows and 50% overlap, and compute the pressure-gradient source from the resulting velocity fields. Re-run the Section 3.1 comparison (SIREN 1x64/20-30, SIREN 1x256/30-30, OS-MODI, GFI) at the same noise levels and report RMAE; also test a variant with noise artificially concentrated near the domain boundaries. If SIREN 1x64/20-30 does not achieve the lowest RMAE at 5–10% added noise, the architecture-based filtering claim fails for realistic, correlated PIV noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SIREN's architecture can be tuned to filter noise from velocimetry data rests on an unverified spectral-separability premise. In Section 3.1 the bandlimit rule ω0 = c·kmin (c ≈ 2–3) is introduced and the denoising benefit is demonstrated by reducing hidden layers/nodes, but the supporting experiments add only isotropic Gaussian velocity noise. Real PIV/LPT error is not white: it is spatially correlated (interrogation-window smoothing, peak locking, bias errors), has low-wavenumber components, and is strongest near boundaries [27, 9, 25]. Such noise overlaps the signal band, so a low-pass SIREN filter would attenuate signal and noise together. The single experimental case (§3.3) is qualitative — the SIREN and OS-MODI traces are similar and the outer solution is explicitly non-physical — so it cannot confirm the filtering mechanism under realistic conditions. If real noise is not frequency-separable from the pressure signal, the headline advantage over OS-MODI erodes toward parity with OS-MODI's inherent regularization (CG/Laplacian smoothing described in the Appendix), and the paper's contribution reduces to mesh-freeness alone, which is still useful but not the advertised noise robustness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a mesh-free pressure reconstruction method based on SIREN, a sinusoidal implicit neural representation, and compares it against OS-MODI and GFI on a two-dimensional slice of the JHTD isotropic turbulence dataset with synthetic Gaussian velocity noise, on structured and unstructured meshes, and on a single experimental PIV wake dataset. The central claims are that SIREN is intrinsically mesh-free, that its architecture can be tuned to filter velocity-derived noise, and that it outperforms OS-MODI and GFI on ill-conditioned unstructured meshes. The authors introduce a frequency-scaling rule omega0 = c*kmin with c in [2,3] and show that shallower networks with smaller omega0 reduce noise sensitivity.","tokens_in":18320,"tokens_out":4775,"duration_ms":46764,"significance":"If validated, the proposed SIREN approach would offer a practically useful mesh-free alternative for pressure integration from PIV/LPT data, particularly in complex geometries where mesh generation is difficult. The benchmark protocol is a strength: all methods share the same source term and are compared directly against DNS ground truth. The spectral analysis in Fig. 4 is informative and suggests that SIREN can resolve substantially higher wavenumbers than the two mesh-based methods. However, the headline noise-robustness claim relies on an unverified spectral-separability premise and is demonstrated only with isotropic Gaussian noise on a single turbulence dataset, with no error bars or repeated realizations. The experimental case is qualitative and the authors explicitly caution that extrapolated SIREN values may be non-physical. The significance is therefore conditional pending stronger validation of the filtering mechanism.","major_comments":[{"comment":"The RMAE curves in Fig. 3 are based on a single realization for each noise level, with no confidence intervals or repeated trials. The conclusion that SIREN 1x64/20-30 \"emerges as the most effective\" at 10% velocity error could be within run-to-run variability of the random noise draw. Without multiple noise realizations or an error analysis, the statistical significance of the difference between SIREN and OS-MODI is not established, and this is load-bearing for the abstract's claim of superior performance in noisy environments.","section":"Section 3.1, Fig. 3"},{"comment":"The noise-filtering claim rests on the premise that noise in the pressure-gradient source term occupies frequencies above the signal band, so that a SIREN with reduced architecture and omega0 = c*kmin filters noise without attenuating signal. The experiments add only isotropic Gaussian velocity noise, which is spectrally white. Real PIV/LPT error is spatially correlated (interrogation-window smoothing, peak locking, boundary bias) and often strongest near boundaries, so it may overlap the signal band. The single experimental case in Section 3.3 shows only qualitative agreement; the authors state that the SIREN outer solution \"should not be relied upon\" and \"may yield non-physical results.\" This is insufficient to confirm the architecture-based filtering mechanism under realistic noise conditions.","section":"Sections 3.1 and 3.3"},{"comment":"The bandlimit rule omega0 = c*kmin with c in [2,3] is calibrated on the same JHTD dataset used for the performance evaluation, and the paper reports only two subdomains of that dataset plus one experimental case. This is a test-set tuning concern: the architecture and scaling choices are selected on the same data used to demonstrate their benefit. A validation on an independent flow case (for example, the Taylor-Green vortex or a different turbulence database) would be needed to show that the rule and the reported noise robustness generalize beyond the tuned case.","section":"Section 3.1, footnote 1"},{"comment":"The unstructured-mesh comparison uses a single random seed for each mesh type (Perturbed and Random). The claimed advantage of SIREN on the random mesh (RMAE 0.0031 versus 0.0213 for OS-MODI and 0.0401 for GFI) might be sensitive to the particular point distribution drawn. Without multiple independent seeds or a perturbed-seed analysis, the robustness of this conclusion is not demonstrated. Since the mesh-free property is one of the two central contributions, this point needs a stronger statistical basis.","section":"Section 3.2"}],"minor_comments":[{"comment":"The loss integral is written as ∫_Ω ||∇xΦ(x) - ∇p(x)|| dx, but Ω is a subset of R² and the integrand is a 2D vector; the notation should use dA or dΩ, and the norm without a subscript could be ambiguous.","section":"Equation (4)"},{"comment":"The phrase \"results (not shown here for brevity)\" for the epoch count is a missing support: either include the plot or state the saturation behavior in one sentence so the reader can assess the convergence criterion.","section":"Section 3.1"},{"comment":"The caption describes \"transparency levels\" for the number of nodes, but the text refers to lighter, intermediate, and darker shades; this is confusing. The legend should be clarified to indicate whether opacity or color is used.","section":"Figure 5"},{"comment":"The acronym RMAE is used without definition before first use; it should be spelled out as relative mean absolute error at its first occurrence.","section":"Section 3.1"},{"comment":"The initialization w0 ∼ U(−1/n, 1/n) uses n without explicitly defining it as the dimension of the input coordinates; this should be stated for clarity.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central noise-filtering claim is currently supported only by synthetic white-noise experiments and is explicitly cautioned against in the experimental section. I would recommend requiring either a realistic correlated-noise test or a second independent DNS validation before publication. The single-realization statistics in Figs. 3 and 7 are also a reproducibility concern. The mesh-free advantage is well demonstrated and is the stronger of the two contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Fernando and Renato have put together a genuinely useful study. The core idea—train a SIREN to match the measured pressure gradient, with the input frequency scaled to the domain's kmin, and use layer/node count as a manual bandlimit—is not in the literature as far as I can tell, and the spectral characterization in Figs. 4 and 5 is the right kind of evidence for why it works. On the structured mesh with added Gaussian noise, SIREN with the smaller 1x64/20-30 config does beat OS-MODI and GFI, and on the random unstructured mesh the improvement (RMAE 0.0031 vs 0.0213/0.0401) is substantial. The benchmark protocol is fair: same source term, direct ground-truth comparison, and the noise is injected into the velocity field before derivatives, not into the pressure gradient itself. That is a real plus.\n\nThe soft spots are the ones the reader flagged, and I largely agree with the stress-test note. The architecture that wins is switched from section to section, tuned with knowledge of the ground truth, and there are no confidence intervals—these are single-realization errors. The RBF meshless method (Sperotto et al.) and a PINN baseline are cited but never run, so the 'mesh-free advantage' claim is not tested against its closest mesh-free competitor. And the experimental case is qualitative; the authors themselves say the SIREN extrapolated solution 'should not be relied upon.' All of that is reportable but not disqualifying.\n\nOn the bigger worry—that real PIV noise is correlated and boundary-dominated, so the spectral-separability premise may fail—I'd say it's a legitimate concern about generalization, but not a flaw in what's shown. The paper never claims to handle non-white noise; it claims the architecture can be tuned to filter noise, and it demonstrates that for the noise model used. The authors are actually transparent about the caveats: the c≈2–3 rule is empirical, the single experimental case is limited, and the extrapolated field is flagged as unreliable. That transparency counts for something.\n\nWho is this for? Anyone doing LPT or PIV pressure reconstruction who is tired of mesh generation. It's a serious candidate for the tracking application, and the spectral analysis will be useful to people tuning SIREN-like networks for other gradient-integration problems. I'd send it to review—it deserves a serious referee—but I'd tell the referee to ask for code/data release and a held-out tuning protocol. As it stands, I wouldn't cite the headline noise-robustness claim without qualification, but I'd cite the mesh-free demonstration and the spectral characterization.","headline":"A useful, well-benchmarked application of SIREN to pressure reconstruction, with real promise for LPT, but the headline noise-filtering claim is demonstrated only on white noise and the paper needs code and held-out tuning before I'd rely on it.","tokens_in":19043,"tokens_out":1988,"would_cite":true,"duration_ms":19961,"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":"This paper claims that a sinusoidal representation network can reconstruct pressure fields from image velocimetry data without a mesh, with architecture choice acting as a tunable noise filter, outperforming OS-MODI and GFI on…","keywords":["pressure reconstruction","SIREN","implicit neural representation","particle image velocimetry","mesh-free methods","pressure gradient integration","noise filtering","unstructured meshes"],"falsifier":"Add correlated, boundary-dominated noise with spectral content inside the signal band to the JHTDB velocity field and compare SIREN RMAE with OS-MODI; if SIREN no longer beats the mesh-based solvers, the architecture-as-filter claim fails. Alternatively, compare SIREN's extrapolated wall pressure against the pressure transducers in the bluff-body experiment; the authors already warn that extrapolated values may be non-physical.","tokens_in":17872,"feed_emoji":"🌊","tokens_out":5984,"duration_ms":57439,"temperature":0.7,"pith_summary":"This paper proposes using a sinusoidal representation network (SIREN) to reconstruct pressure fields from image velocimetry data by integrating the pressure gradient directly, without building a mesh. The authors argue that because the network maps coordinates to pressure through sine activations, it can be initialized and sized so that its frequency bandlimit acts as a filter, removing noise from the measured velocity field while preserving the true pressure signal. On a randomly seeded unstructured mesh, the method reports a relative mean absolute error of 0.0031 versus 0.0213 for one-shot matrix omnidirectional integration and 0.0401 for Green's function integration, and it resolves spectral content roughly an order of magnitude beyond those methods. If correct, the approach turns architecture choice into a denoising knob and removes mesh quality as a source of error for particle tracking and complex geometries.","feed_headline":"SIREN rebuilds pressure fields from noisy flow images, no mesh needed","feed_subtitle":"Architecture tuning acts as a bandpass filter, beating grid-based solvers on unstructured data.","key_machinery":"The load-bearing object is the SIREN: a fully connected network with sine activation functions, trained with automatic differentiation so that $\\nabla_x \\Phi(x)$ approximates the measured pressure gradient from $\\nabla p = -\\rho Du/Dt + \\mu \\nabla^2 u$. Its frequency behavior follows a \"frequency factoring\" mechanism: the first layer acts as a Fourier feature map, and each sinusoidal neuron generates integer linear combinations of input frequencies with bounded amplitudes, so the network's spectral content is set at initialization by the input scaling $\\omega_0 = c k_{\\min}$ with $2 \\lesssim c \\lesssim 3$ and the hidden-layer scalings $\\omega_i \\approx 20$-$30$. This is what lets the same network act as an integrator and as a tunable bandpass filter.","core_discovery":"The central claim is that a SIREN, trained so that the gradient of its output matches the pressure gradient implied by the momentum equation, is a mesh-free pressure integrator whose architecture controls the trade-off between resolution and noise sensitivity. The paper shows that the input frequency scale can be set from the smallest wavenumber of the domain, and that fewer hidden layers and nodes suppress high-frequency noise while deeper or wider networks resolve finer scales. In noise-free structured-mesh cases the method matches OS-MODI; under added velocity noise the shallow 1x64/20-30 network has the lowest error of all tested methods; on an ill-conditioned random mesh it is several times more accurate than OS-MODI and GFI. The authors present the bluff-body experiment as a qualitative demonstration only, and explicitly warn that extrapolated outer solutions should not be relied upon.","pith_inferences":["Editorial extension: if real PIV noise is broad-band and concentrated at walls rather than spectrally separated from the signal, the architecture-as-filter advantage could shrink; a test with correlated noise would settle the question.","Editorial extension: the same frequency-scaling rule could apply to other inverse problems that integrate a measured gradient, such as force or heat-flux reconstruction from particle data.","Editorial extension: the $c$ range of 2-3 for $\\omega_0$ is empirical, so a data-driven selection of the bandlimit from validation error would make the method more portable.","Editorial extension: volumetric Lagrangian particle tracking with voids is the natural next case, since the mesh-free property should matter most where connectivity is absent."],"forward_implications":["Users can integrate pressure from scattered particle-tracking or PIV points with no mesh generation, sidestepping ill-conditioned cells entirely.","Choosing a shallower or narrower SIREN trades high-wavenumber detail for noise rejection, giving the experimenter a single lever to tune denoising.","In the noiseless spectral test the SIREN resolves wavenumbers roughly ten times higher than OS-MODI and GFI, so small-scale pressure features become accessible.","Extrapolation outside the measured region is unreliable and may return non-physical values, so the method should be restricted to the data domain.","The method's training cost (about 16 minutes for the deep configuration on the benchmark) is moderate, though the authors do not claim a head-to-head optimized comparison."],"supporting_citations":[{"why":"Supplies the SIREN architecture and its initialization scheme.","marker":"[35]"},{"why":"Gives the frequency-factoring mechanism used to justify bandlimit control.","marker":"[26]"},{"why":"Provides the OS-MODI baseline method and its matrix-inversion formulation.","marker":"[46]"},{"why":"Provides the Green's function integral baseline and its discretized convolution.","marker":"[41]"},{"why":"Establishes how boundary and domain error propagates in pressure reconstruction, motivating the noise study.","marker":"[27]"},{"why":"Supplies the isotropic1024coarse DNS dataset with ground-truth pressure used for quantitative tests.","marker":"[16]"},{"why":"Supplies the slanted-bluff-body PIV experiment and surface pressure data for the qualitative demonstration.","marker":"[47]"},{"why":"Provides benchmark assessment context and validation cases for PIV-based pressure methods.","marker":"[6]"}],"fun_headline_variants":["SIREN pressure fields beat mesh solvers in noisy, unstructured flows","Mesh-free SIREN reconstructs pressure from noisy velocimetry","Neural pressure reconstruction: SIREN filters noise without meshes","SIREN pressure integration handles noise, no grid connectivity","Architecture-tuned SIREN outperforms grid-based pressure methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The noise-robustness claim assumes that velocity-derived pressure-gradient noise lives at higher frequencies than the true pressure signal, so shrinking the network's bandlimit removes noise without removing signal.","fun_headline_variants_meta":{"raw":{"variants":["SIREN pressure fields beat mesh solvers in noisy, unstructured flows","Mesh-free SIREN reconstructs pressure from noisy velocimetry","Neural pressure reconstruction: SIREN filters noise without meshes","SIREN pressure integration handles noise, no grid connectivity","Architecture-tuned SIREN outperforms grid-based pressure methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000467,"raw_usage":{"total_tokens":2336,"prompt_tokens":960,"completion_tokens":1376,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":1287}},"tokens_in":576,"tokens_out":1376,"duration_ms":10593,"temperature":1.0,"reasoning_tokens":1287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:28:09.135965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Add correlated, boundary-dominated noise with spectral content inside the signal band to the JHTDB velocity field and compare SIREN RMAE with OS-MODI; if SIREN no longer beats the mesh-based solvers, the architecture-as-filter claim fails. Alternatively, compare SIREN's extrapolated wall pressure against the pressure transducers in the bluff-body experiment; the authors already warn that extrapolated values may be non-physical.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SIREN architecture and its initialization scheme."},{"cited_title":"Taming the frequency factory of sinusoidal net- works, 2024","cited_arxiv_id":null,"evidence_quote":"Gives the frequency-factoring mechanism used to justify bandlimit control."},{"cited_title":"Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T","cited_arxiv_id":null,"evidence_quote":"Provides the Green's function integral baseline and its discretized convolution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes how boundary and domain error propagates in pressure reconstruction, motivating the noise study."},{"cited_title":"A public tur- bulence database cluster and applications to study 14 Renato F","cited_arxiv_id":null,"evidence_quote":"Supplies the isotropic1024coarse DNS dataset with ground-truth pressure used for quantitative tests."},{"cited_title":"Reynolds number and slant angle effects on the flow over a slanted cylinder afterbody","cited_arxiv_id":null,"evidence_quote":"Supplies the slanted-bluff-body PIV experiment and surface pressure data for the qualitative demonstration."},{"cited_title":"Charonko, Cameron V","cited_arxiv_id":null,"evidence_quote":"Provides benchmark assessment context and validation cases for PIV-based pressure methods."}],"review_version":1}