{"id":"b10f24d3-0ca9-4767-a12c-d631115e89cb","arxiv_id":"2507.16431","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A physics-informed convolutional neural operator predicts scattered Helmholtz wavefields with up to 53% lower relative error than its purely data-driven counterpart on held-out velocity models.","lead":"The authors add a physics-based penalty term to a convolutional neural operator that predicts frequency-domain seismic wavefields, and they report that it beats the data-only version on unseen velocity models, especially at higher frequencies. The result matters because faster wavefield surrogates could lower the cost of seismic imaging and full-waveform inversion, though the gains are shown only on small 2D synthetic problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 8's PDE residual is computed over the full grid with the singular Hankel background U0 included at the source; without a stated mask or taper, the physics loss is dominated by an unrepresentable singularity, so the claimed 53% improvement may be an artifact rather than evidence of…","rationale":"The controlled architecture comparison and the authors' own discussion of instability and loss-weight sensitivity are read in good faith; they do not by themselves defeat the empirical claim. The single load-bearing condition for the central claim is that Eq. 8 behaves as a Helmholtz residual over the wavefield. That condition is broken, or at least unverified, at the source because U0 is log-singular (Eq. 3) and the residual is evaluated pointwise over the whole grid (Section 2.3) with no mask or taper stated. Since the CNO output is band-limited, the residual cannot be small near the source, so its gradient is concentrated there; this directly explains the reported training instability and makes the physics loss's contribution to generalization uncontrolled. The proposed masked-retraining test is decisive: if the PICNO advantage persists with the singular region excluded, the mechanism is supported; if not, the headline is an artifact. Other concerns such as missing seed variance, a single OOD example, the uncited CNO baseline, and lambda sensitivity are real but secondary because they affect confidence in the magnitude of the gain, not the causal story. The reader's weakest assumption already names the source singularity, so agreement is agree. The verdict remains CONDITIONAL, pending the source-mask clarification.","tokens_in":15200,"tokens_out":7849,"duration_ms":85881,"concrete_test":"Retrain PICNO with the PDE residual masked or smoothly tapered within a radius of 3-5 grid cells around the source, equivalently excluding the singular U0 term from Eq. 8 in that region, keeping all other hyperparameters identical. If Table 1's 12 Hz error stays near 0.23 and the qualitative field plots hold, the physics constraint is the cause; if the error reverts toward CNO's 0.50, the claimed advantage is attributable to the unresolved source-singularity term. As a diagnostic, also record the spatial distribution of the squared residual on the trained model: if the top 1% of grid points centered on the source account for most of the norm, the unmasked loss is dominated by the singularity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adding the Helmholtz residual (Eq. 8) to the data loss (Eq. 7) regularizes the operator and improves test generalization (Table 1). For that claim to hold, the residual must actually constrain the network output toward solutions of the scattered Helmholtz equation over the wavefield. This condition is least secure at the source. U0 in Eq. 3 is the Hankel function H0^(2)(kr), which behaves as (2i/pi) log r near the source. Eq. 8 is stated to be evaluated pointwise over the input grid using an eighth-order finite-difference stencil, and no masking, tapering, or source-region exclusion is mentioned anywhere in Section 2.3, Section 3, or the Discussion. The network output, by contrast, is a band-limited convolutional field (CNO low-pass filters, Section 2.2); it cannot reproduce a logarithmic singularity or absorb the singular forcing term omega^2 (1/v^2 - 1/v0^2) U0. The residual near the source therefore cannot be made small in the norm of Eq. 8, and its gradient is dominated by a few grid points. This matches the admitted abrupt fluctuations and outliers in Section 3.1 and the sensitivity to lambda in Section 4.1. If the source region is not masked, the physics loss is not a faithful Helmholtz constraint; it is a source-dependent penalty whose effect on the learned operator is uncontrolled. The observed Table 1 improvement (0.50 to 0.23 at 12 Hz) could then come from this artifact rather than from enforcing the PDE, which is the paper's central causal story. This is not a claim of dishonesty; it is an omitted implementation detail that determines whether the headline result supports the stated mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript introduces the physics-informed convolutional neural operator (PICNO), which predicts frequency-domain scattered wavefields for the Helmholtz equation by combining a convolutional neural operator (CNO) with a PDE-residual loss. The inputs are the analytic background wavefield U0 and a velocity model; the output is the scattered wavefield δU. Experiments on synthetic datasets derived from OpenFWI at 8, 10, and 12 Hz with fixed-depth sources (Case 1), at 12 and 14 Hz with randomly distributed source locations (Case 2), and a single out-of-distribution velocity model with faults are presented. The central claim is that the physics term improves generalization, with relative L2 errors dropping from 0.50 to 0.23 at 12 Hz (a 53.1% relative reduction) and qualitatively better wavefield detail.","tokens_in":15524,"tokens_out":8323,"duration_ms":80188,"significance":"The problem is well-motivated for seismic imaging and inversion. The experimental design is sound in principle: the comparison is controlled (same architecture, only the physics loss term is added), the improvements are consistent across several frequencies and both source-distribution cases, and the code is made publicly available via GitHub. If the reported gains are robust, the paper would provide a valuable demonstration that physics-informed regularization can improve the generalization of neural operators in data-sparse geophysical settings. However, the current evidence is weakened by an unstated treatment of the source singularity in the PDE loss, the lack of multiple-seed statistics, and the anecdotal nature of the out-of-distribution evaluation. These are load-bearing issues, but they are addressable within the scope of a revision.","major_comments":[{"comment":"The PDE residual in Eq. (8) is stated to be evaluated pointwise over the input grid using an eighth-order finite-difference stencil, and the background field U0 in Eq. (3) is the Hankel function H0^(2), which is logarithmically singular at the source. The residual therefore includes the singular term ω^2(1/v^2 − 1/v0^2)U0 plus the Laplacian of the predicted field, and the paper never states that the source region is masked, tapered, or excluded. Because the CNO output is a smooth, band-limited convolutional field (Section 2.2), the residual near the source cannot be made small, and the gradient of Jpde is plausibly dominated by a few grid points. This makes the physics loss an uncontrolled source-dependent penalty rather than a faithful Helmholtz constraint, and it weakens the causal interpretation of the improvements in Table 1. Please specify whether a source mask or taper is used, and if not, either introduce one, modify the loss, or provide evidence that the source-region contribution does not dominate the residual.","section":"2.3, Eq. (8)"},{"comment":"The headline results in Table 1 are single runs without variance information. The authors themselves document abrupt fluctuations and outliers in the validation error (Section 3.1, Figure 3(d)–(f)) and a sensitive dependence on the loss weight λ (Section 4.1). Without multiple random seeds or equivalent statistical treatment, the reported relative improvements of 27.4%, 35.2%, and 53.1% are not established as reproducible. Please report mean and standard deviation over at least three independent training runs (ideally five) for both CNO and PICNO at each frequency and case, and state the exact λ values used for every configuration. If the advantage does not persist across seeds, the claims should be qualified accordingly.","section":"3, Table 1"},{"comment":"The out-of-distribution evaluation is based on a single visual example (Figure 10) with no quantitative error metric. To substantiate the claim that PICNO generalizes to faulted and sharp-discontinuity models, report the relative L2 error (or a similar metric) for the OOD model for both CNO and PICNO, and ideally for a small set of OOD models. Without numbers, the OOD conclusion is anecdotal and does not support the strong statement in the abstract and conclusions.","section":"3.3, Figure 10"},{"comment":"The λ-sensitivity experiment is described only through the figure caption, which labels 'PICNO case1' and 'PICNO case2' with loss ratios of 10:1 and 8:1, but the text does not specify which case or frequency this refers to, nor the exact λ values used in the main experiments. Since the authors state that λ is selected empirically and that no universal ratio exists, the chosen λ for each training configuration (Cases 1 and 2, each frequency) must be reported, and the authors should justify that the headline results are not a product of selecting λ on the validation set. Ideally, also provide a sensitivity analysis over a range of λ values for at least one representative configuration.","section":"4.1, Figure 11"}],"minor_comments":[{"comment":"The section title contains a typo: 'F ormulation' should be 'Formulation'.","section":"2.3"},{"comment":"Check the argument of the Hankel function in Eq. (3); the wavenumber should be ω/v0 rather than ω/v0^2 as printed. If this is a typesetting error, correct it; if it reflects the implementation, the physics loss would use an incorrect background wavefield.","section":"Eq. (3)"},{"comment":"The captions contain garbled path strings such as '/gid00051/gid00068/...'; these are likely LaTeX artifacts and should be removed.","section":"Figures 4–10"},{"comment":"The opening sentence of Section 4 refers to 'the Example section'; the intended term is 'Results' section.","section":"4"},{"comment":"The architecture description could be clearer; in particular, 'U-Net-style patching' and the exact composition of the D, H, I, and R blocks are hard to follow. A pseudocode or a table summarizing layer configurations would aid reproducibility.","section":"2.2"},{"comment":"The sentence 'We show the the relative error' contains a duplicated definite article.","section":"3.1"}],"recommendation":"major_revision","confidential_remarks":"This paper is within the scope of JGR: Machine Learning and Computation and the central idea is appealing. My recommendation of major revision is driven by three load-bearing gaps: the unspecified handling of the source singularity in the PDE loss, the absence of multi-seed error bars for the headline numbers, and the anecdotal OOD result. All three are fixable with additional experiments and reporting. The authors should also be asked to report the exact λ values used for each table entry. If these are addressed in a revision, I would expect the paper to be a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a well-posed, incremental methods paper. The key experiment—same CNO architecture trained with and without the Helmholtz residual—shows consistent error reductions, and the authors are upfront about training instability and lambda sensitivity. But there is a likely load-bearing omission: the PDE residual in Eq. (8) is evaluated with the singular Hankel background field U0 at the source, with no stated mask or taper. That could explain the abrupt loss fluctuations and makes the causal interpretation of the gains less secure.\n\nWhat is new: PINO losses and the scattered-field input are established, but the specific PICNO combination is not in prior work. The controlled comparison isolates the physics term, which is the right thing to do. The data augmentation with smoothed velocities is a sensible way to stretch a small dataset. The paper also provides a code link and discusses limitations honestly.\n\nSoft spots, in order of severity: (1) The source singularity in Eq. (8) is never addressed. The network output is band-limited, so it cannot represent the logarithmic singularity in the true scattered field when the velocity at the source differs from v0. The PDE residual will be dominated by a few near-source points, and the weight lambda is tuned around that. This needs a mask or a careful handling statement. (2) The headline numbers are single runs without seed variance; the 53% improvement at 12 Hz is one realization. (3) The OOD test is a single qualitative example. (4) The original CNO paper is not cited, which is an odd omission given how central the architecture is. None of these are fatal, but they mean the claim that the physics term regularizes generalization is not yet fully pinned down.\n\nWho this is for: readers in geophysical ML who want a comparison of physics-informed vs data-only operator learning on a realistic frequency-domain problem. It deserves a serious referee. I would ask for the source-mask detail, multiple seeds, and a quantitative OOD evaluation before accepting.\n\nNet: promising, plausible, but the mechanism is under-specified. Recommended for peer review with major revisions.","headline":"A modest but well-controlled PINO paper whose central causal claim is undercut by an unaddressed singularity in the PDE loss.","tokens_in":16117,"tokens_out":3613,"would_cite":false,"duration_ms":41757,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65N06","86A15"],"pacs":["91.30.-f","02.60.-x","07.05.Mh"],"model":"deepseek-v4-flash","headline":"Adding the Helmholtz residual to a neural operator's loss cuts wavefield error by up to 53.1%.","keywords":["physics-informed neural operator","Helmholtz equation","seismic wavefield prediction","convolutional neural operator","scattered wavefield","generalization","full waveform inversion","frequency-domain wavefield"],"falsifier":"Retrain PICNO with the source neighborhood removed from the physics loss (and the same hyperparameters), then compare test errors; if the 53.1% improvement over CNO shrinks sharply, the claimed regularization is coming from the singular source region rather than from honoring the wave equation globally. A second check is to evaluate the PDE residual of PICNO's predictions on out-of-distribution models; if that residual is not systematically smaller than CNO's, the physics constraint is not being enforced where it is claimed to matter.","tokens_in":14955,"feed_emoji":"🌊","tokens_out":7635,"duration_ms":68488,"temperature":0.7,"pith_summary":"This paper tries to establish that a convolutional neural operator can learn the frequency-domain Helmholtz solution operator from a small training set if the Helmholtz residual is added to the training loss, because the physics constraint acts as a regularizer. The network maps the analytically computed background wavefield and the velocity model to the scattered wavefield, and the physics loss penalizes violations of the scattering-form wave equation pointwise with an eighth-order finite-difference Laplacian. The headline evidence is that on held-out velocity models the relative L2 error drops from 0.50 to 0.23 at 12 Hz (a 53.1% relative reduction), with smaller gains at 8 and 10 Hz. The paper further reports that PICNO preserves wavefront detail at high frequency and on out-of-distribution fault models, while the purely data-driven CNO shows phase mismatch and amplitude damping. If the claim is right, fast wavefield surrogates for seismic imaging and waveform inversion do not necessarily require massive training datasets.","feed_headline":"Physics loss cuts neural-operator wavefield error by 53%","feed_subtitle":"At 12 Hz, test error on unseen velocity models drops from 0.50 to 0.23 with sharper wavefronts.","key_machinery":"The load-bearing object is the physics loss $J_{\\mathrm{pde}}$: the squared residual of the scattering-form Helmholtz equation, evaluated pointwise with an eighth-order finite-difference Laplacian over the entire grid. Substituting the network output $G_\\theta$ into this residual gives gradients that push the prediction toward satisfying the wave equation, not merely matching labels. The second load-bearing object is the analytically computed background wavefield $U_0$, a Hankel function of the second kind for a constant background, which encodes source position and frequency as two input channels and defines the target as the scattered field $\\delta U = U - U_0$.","core_discovery":"The central claim is that physics-informed training changes what the operator learns, not just how well it fits. PICNO uses the same convolutional neural operator architecture as the data-driven baseline, with convolutions performed in physical space, and it takes as input the real and imaginary parts of the background wavefield $U_0$ plus the velocity model $v$, producing the real and imaginary parts of the scattered wavefield $\\delta U$. The physics loss $J_{\\mathrm{pde}}$ is the squared residual of the scattering-form Helmholtz equation, $\\frac{\\omega^2}{v^2}G_\\theta + \\nabla^2 G_\\theta + \\omega^2(\\frac{1}{v^2}-\\frac{1}{v_0^2})U_0$, evaluated pointwise on the whole grid with an eighth-order finite-difference stencil. The paper reports that, with 3600 training samples per frequency, PICNO achieves relative L2 errors of 0.18, 0.22, and 0.23 at 8, 10, and 12 Hz, versus 0.25, 0.34, and 0.50 for CNO, and that the gap persists when sources are randomly placed throughout the model and on an out-of-distribution fault-model test. The authors also state that the PDE term destabilizes training, producing abrupt validation fluctuations and occasional jumps in the loss.","pith_inferences":["The residual $J_{\\mathrm{pde}}$ is evaluated pointwise over the whole grid with no stated masking or down-weighting of the source neighborhood, where $U_0$ has a logarithmic singularity; if the source region is not excluded, part of the reported gain may reflect fitting that singular structure rather than a global wave-equation constraint.","A clean test of the regularization story would be to retrain with the residual computed only at points at least one wavelength from the source; if the 53.1% improvement vanishes, the source-region term is doing most of the work.","If the effect is a genuine inductive bias, the same scattered-field-plus-PDE-loss recipe should transfer to other elliptic equations with analytic background solutions, such as the damped Helmholtz equation, where the operator only has to learn the correction field.","The reported 0.83-second inference time does not include the extra cost of evaluating the PDE residual during training, so the speed claim applies to inference-only use of the trained surrogate rather than to end-to-end inversion loops."],"forward_implications":["On the 800-sample test set, PICNO's relative L2 error is 0.18, 0.22, and 0.23 at 8, 10, and 12 Hz against CNO's 0.25, 0.34, and 0.50, a 27.4%, 35.2%, and 53.1% relative reduction.","PICNO predicts a frequency-domain wavefield in about 0.83 seconds versus about 3 seconds for the numerical solver, so accuracy gains do not sacrifice the speed advantage of neural operators.","Training on 72 base velocity models (3600 samples per frequency) is enough for PICNO to generalize to held-out velocity models, indicating that the PDE constraint reduces the data requirement.","PICNO also generalizes to source locations distributed throughout the domain and to an out-of-distribution velocity model with faults, where the data-only CNO shows much larger errors.","The physics loss introduces training instability, with abrupt validation-error fluctuations and occasional loss jumps, so the improved generalization comes with a harder optimization landscape."],"supporting_citations":[{"why":"Provides the purely data-driven neural-operator baseline and the background-wavefield input parametrization that PICNO extends with a physics loss.","marker":"Huang & Alkhalifah (2025)"},{"why":"Supplies the physics-informed neural operator methodology of adding a discrete PDE residual to the data loss.","marker":"Li et al. (2024)"},{"why":"The source of the physics-constrained loss idea from PINNs that the paper transfers to operator training.","marker":"Raissi et al. (2019)"},{"why":"Introduces the scattered and background decomposition of the Helmholtz equation used to define the learning target and the residual in the physics loss.","marker":"Alkhalifah et al. (2021)"},{"why":"Supplies the velocity-model benchmark from which the training, validation, and test samples are drawn.","marker":"Deng et al. (2022)"},{"why":"Shows neural operators can be far faster than numerical solvers for seismic wavefield modeling, the speed motivation PICNO inherits.","marker":"Yang et al. (2023)"}],"fun_headline_variants":["Physics loss cuts wavefield error 54% at 12 Hz","PDE-informed operator halves wavefield error at 12 Hz","Physics-informed operator beats data-only wavefield prediction","Physics loss sharpens high-frequency wavefield predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that evaluating the Helmholtz residual pointwise over the whole grid—including grid points right next to the source, where the analytic background field blows up logarithmically—gives a stable and physically meaningful training signal; the paper does not state that the source region is masked or down-weighted.","fun_headline_variants_meta":{"raw":{"variants":["Physics loss cuts wavefield error 54% at 12 Hz","PDE-informed operator halves wavefield error at 12 Hz","Physics-informed operator beats data-only wavefield prediction","Physics loss sharpens high-frequency wavefield predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001276,"raw_usage":{"total_tokens":5232,"prompt_tokens":975,"completion_tokens":4257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":4191}},"tokens_in":591,"tokens_out":4257,"duration_ms":36132,"temperature":1.0,"reasoning_tokens":4191,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:09:24.742495+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain PICNO with the source neighborhood removed from the physics loss (and the same hyperparameters), then compare test errors; if the 53.1% improvement over CNO shrinks sharply, the claimed regularization is coming from the singular source region rather than from honoring the wave equation globally. A second check is to evaluate the PDE residual of PICNO's predictions on out-of-distribution models; if that residual is not systematically smaller than CNO's, the physics constraint is not being enforced where it is claimed to matter.","supporting_citations":[{"cited_title":"\\ Alkhalifah, T","cited_arxiv_id":null,"evidence_quote":"Provides the purely data-driven neural-operator baseline and the background-wavefield input parametrization that PICNO extends with a physics loss."},{"cited_title":", Perdikaris, P","cited_arxiv_id":null,"evidence_quote":"The source of the physics-constrained loss idea from PINNs that the paper transfers to operator training."},{"cited_title":", Song, C","cited_arxiv_id":null,"evidence_quote":"Introduces the scattered and background decomposition of the Helmholtz equation used to define the learning target and the residual in the physics loss."},{"cited_title":", Feng, S","cited_arxiv_id":null,"evidence_quote":"Supplies the velocity-model benchmark from which the training, validation, and test samples are drawn."},{"cited_title":", Gao, A F","cited_arxiv_id":null,"evidence_quote":"Shows neural operators can be far faster than numerical solvers for seismic wavefield modeling, the speed motivation PICNO inherits."}],"review_version":1}