{"id":"0ac0d07f-8363-4c08-9a5a-6dc0fc6923a1","arxiv_id":"2508.10555","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"A continuous neural-field parameterization of the contrast source gives an unsupervised, differentiable inversion framework that improves CSI accuracy and enables super-resolution.","lead":"This paper replaces the usual pixel-by-pixel contrast source in inverse scattering with a small neural network that predicts the source continuously across space and transmitter settings. The result is a single differentiable, label-free framework for full and phaseless scattered-field data, reporting better accuracy and robustness than conventional contrast source inversion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Numerical gains and super-resolution may stem from same-model inverse crime and unconstrained interpolation; independent forward-model validation is required.","rationale":"The reader's weakest assumption was that the VIE accurately models the data-generation process. I agree that this is the key unvalidated premise, and I sharpen it to a concrete inverse-crime risk: if the same discretized VIE is used both to synthesize data and to invert, the reported accuracy and robustness are upper bounds that do not transfer to real measurements. I add a second, distinct concern about the super-resolution claim: querying a continuous network at finer coordinates does not by itself add information unless the forward model is enforced at those coordinates, so fine-grid data/state residuals must be reported to rule out pure interpolation. The supplied full text is largely non-decodable, so this assessment is based on the abstract and readable equation fragments; that also supports keeping the reader's UNVERDICTED verdict rather than moving to acceptance or rejection. The concern is significant enough to require an independent forward-model test, but it does not demonstrate that the method is false, only that its central evidence is currently incomplete.","tokens_in":20238,"tokens_out":6308,"duration_ms":77229,"concrete_test":"Run the benchmark with scattered fields generated by an independent forward solver (e.g., FDTD or a VIE discretized on a different grid/quadrature) for identical phantoms and sensor configurations; re-run the proposed method and baseline CSI. If the accuracy/noise-robustness advantage disappears or materially shrinks, the VIE-consistency assumption is load-bearing. In the same run, after coarse-grid optimization, evaluate the full objective (data misfit + state residual) on the claimed fine reconstruction grid and compare against the coarse solution interpolated to that grid; if the fine-grid residual is not lower, the super-resolution claim is interpolation rather than data-driven inference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy and robustness claims are established only if the VIE used inside the objective is an unbiased model of the measurements. The abstract reports numerical experiments, and the visible text shows a VIE-based objective with data and state residuals, but there is no indication of experimental or independent-solver validation. If the synthetic data are generated with the same discretized VIE and the same Green's function as the inversion, the comparison suffers from an inverse crime: the method is tested against the exact forward model it assumes, so discretization error and modeling error are cancelled rather than learned. This would inflate both the accuracy margins over conventional CSI and the apparent noise robustness. The super-resolution claim is separately load-bearing: because the ResMLP is queried at fine coordinates after optimization, the denser points are not constrained by the data equation unless the VIE is re-evaluated there. Without reporting fine-grid data and state residuals, 'super-resolution' may be just the network's implicit smooth interpolation, not new information. The paper's wording that resolution is decoupled from inversion cost is therefore not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a differentiable contrast source inversion (CSI) framework in which the contrast source is represented as a continuous neural field (ResMLP) conditioned on spatial coordinates and transmitter settings. The inversion objective combines the volume integral equation (state equation), the data equation, and total-variation regularization, and the network parameters together with the medium contrast are optimized by automatic differentiation. The framework is unsupervised in the sense of not using ground-truth labels, and it handles full and phaseless data by changing only the data misfit. Numerical experiments reportedly show improved accuracy and robustness over conventional CSI across noise levels and measurement settings, and the continuous parameterization is claimed to enable super-resolution inference at grids finer than the training grid. Ablation studies and comparisons with alternative neural architectures are reported to support the claim that both the contrast-source parameterization and the VIE-based formulation are essential.","tokens_in":20546,"tokens_out":6155,"duration_ms":76586,"significance":"If the performance and super-resolution claims hold, this is a useful contribution: it offers a label-free neural implicit representation that decouples reconstruction cost from output resolution, with a single VIE-constrained objective for full and phaseless data. The problem is timely, and the experimental design in the abstract—noise levels, measurement settings, ablations, and architecture comparisons—is appropriate in principle. The claims are concrete and numerically testable. The main reservations concern the evidence base: the provided full text is only partially decodable, and the visible portions do not establish that the synthetic data are generated independently of the inversion model, nor that the reported super-resolution is constrained by measurements rather than by network smoothness.","major_comments":[{"comment":"The central accuracy and robustness claims rest on the synthetic experiments, but the decipherable text does not state whether the measured data were generated with an independent forward solver/discretization or with the same discretized VIE and Green's function used inside the inversion objective. If the latter, the comparison with conventional CSI suffers from an inverse crime: modeling and discretization errors are canceled, inflating both the accuracy margins and the apparent noise robustness. Please specify the forward model, the Green's function discretization, the mesh used for data generation versus inversion, and whether the CSI baseline uses the same forward model. A test with an independent solver or with experimental data is needed to support the robustness claim.","section":"Numerical experiments (data generation and baselines)"},{"comment":"The ResMLP can be evaluated at arbitrary coordinates, but the optimization constrains the network only at the training-grid points through the state and data equations. If the reported super-resolution is obtained by a final fine-grid evaluation without re-evaluating the VIE and data equation on the fine grid, then the additional pixels are not informed by the measurements; they are the network's implicit interpolation. Please report fine-grid data and state residuals, and compare against a conventional interpolation of the coarse reconstruction. Without this, the statement that super-resolution is achieved and that inversion cost is decoupled from reconstruction fidelity is not established.","section":"Super-resolution claim (Abstract and results section)"},{"comment":"The abstract's final claim that the contrast-source parameterization and VIE formulation are 'both essential' requires the ablations to control for model capacity, optimizer settings, initialization, and regularization weights. The visible portion of the manuscript does not show such controls, nor does it report error bars or the number of independent runs for the tables. If the TV weight and the relative weighting of state-equation and data-equation losses are tuned separately for each noise level and for the baselines, the reported noise robustness and the attribution of gains to specific components may be partly post-hoc. Please provide the full experimental setup, including all hyperparameters, and report variability across runs.","section":"Experimental protocol and ablations"}],"minor_comments":[{"comment":"Define all symbols in the VIE equations precisely before first use. In the decipherable fragments, the contrast function, contrast source, incident field, and scattered field appear with inconsistent transmitter/data indices, and the domains of the operators are not stated.","section":"Notation and equations"},{"comment":"The captions of the figures and tables are hard to follow in the provided version. Please ensure each table reports the metric definition, number of runs, and error bars or confidence intervals, and that each figure caption states the noise level and measurement configuration.","section":"Figures and tables"},{"comment":"The term 'unsupervised' may be misleading: the method is label-free, but it is not trained on a dataset in the usual learning sense. Consider 'physics-driven' or 'self-supervised by the data equation' to avoid confusion with data-driven unsupervised learning.","section":"Terminology"},{"comment":"The 'conventional CSI' baseline should be described in enough detail to be reproducible: update scheme, regularization, stopping criterion, and whether the same TV regularization is used. Without this, the reported improvements are difficult to interpret.","section":"Baseline description"}],"recommendation":"major_revision","confidential_remarks":"The supplied full text is heavily garbled and only partially decodable; I could not verify the equations, tables, and experimental details in full. My assessment is therefore based on the abstract and the decipherable fragments. The stress-test concern about inverse crime is real and should be addressed explicitly in the revision; the super-resolution claim also needs fine-grid residual reporting. If the garbling is a submission defect, please obtain a clean version before the next decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick note before you spend time on it: the body text I received is garbled — Unicode replacement characters throughout — so I can't check the equations, figures, tables, or error bars. This letter is based on the abstract plus visible structure. My honest take: this is a real new combination, not a paper with nothing under the hood. Neural implicit fields have been used in imaging, and CSI is a standard inverse-scattering workhorse; putting the contrast source into a lightweight ResMLP and jointly optimizing network weights and contrast against the VIE state and data equations is a genuine extension, and the phaseless variant is a natural fit. The abstract reports comparisons across noise levels, measurement settings, and ablations against alternative architectures. That's the right kind of evidence for the central claim.\n\nSoft spots: the two things I cannot verify because of the unreadable body. First, the synthetic-data generation. If the measurements are produced with the same discretized VIE and same Green's function the inversion assumes, you get an inverse crime and the accuracy margins are inflated. The stress-test note makes this point, and nothing in the abstract rules it out. Second, the 'super-resolution inference at resolutions finer than the training grid' is the weakest claim. Querying a continuous MLP at denser coordinates gives you a denser field, but not new information unless the data and state equations are re-evaluated at those points. Without reported fine-grid residuals, 'decoupling inversion cost from reconstruction fidelity' is likely just cheap interpolation. That is a real limitation of the claim as stated.\n\nThe full-text problem also means I can't tell whether the authors already address these checks. Maybe the arXiv version is fine and our copy is corrupted; if you can retrieve a readable PDF, look at the numerical setup and the super-resolution experiment. The needed validation is clear: test against an independent forward solver that differs in discretization or Green's function implementation, and report fine-grid data and state residuals after the super-resolution step.\n\nFor the audience: electromagnetic/microwave inverse scattering and ML-for-imaging people will get value from this. It deserves a serious referee—the idea is sensible, the claims are concrete, and the required checks are unambiguous. I'd send it out with a request for code/data release and independent-solver validation. If those hold, I'd be glad to cite it; right now, I wouldn't.","headline":"A genuine CSI + neural-implicit combination that deserves referee time, but the unreadable full text and a likely over-reach on super-resolution mean I'd hold off citing until validated.","tokens_in":20954,"tokens_out":2863,"would_cite":false,"duration_ms":33999,"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 continuous neural-field representation of the contrast source turns inverse scattering into a differentiable, unsupervised optimization that also supports super-resolution.","keywords":["inverse scattering","contrast source inversion","neural implicit representation","phaseless data inversion","volume integral equation","automatic differentiation","total variation regularization","super-resolution"],"falsifier":"Take calibrated experimental scattered-field data from a dielectric phantom with independently known ground-truth contrast, add controlled noise, and run the same inversion. The central claim fails if reconstruction error degrades much faster with noise than in the reported simulations, or if sampling the trained neural field on a grid finer than the training grid produces no visible resolution gain.","tokens_in":20185,"feed_emoji":"📡","tokens_out":6194,"duration_ms":69114,"temperature":0.7,"pith_summary":"The paper tries to establish that inverse scattering can be recast as a fully differentiable, unsupervised optimization by representing the unknown contrast source as a continuous neural field, not as the usual pixel-wise array. A lightweight residual multilayer perceptron maps spatial coordinates and transmitter settings to the contrast source, and the volume-integral-equation state equation, the data equation, and a total-variation regularity term are combined into a single differentiable objective. Network parameters and the medium contrast are then optimized jointly through automatic differentiation, and both full-data and phaseless-data inversion are handled by changing only the data misfit term. The authors report that this scheme improves reconstruction accuracy and noise robustness over conventional contrast source inversion across various noise levels and measurement settings, and that the continuous field can be sampled at resolutions finer than the training grid, so output fidelity no longer dictates the inversion cost.","feed_headline":"A neural field beats pixel grids for inverse scattering","feed_subtitle":"Unsupervised neural inversion boosts accuracy and resolution in scattering reconstruction","key_machinery":"The central object is the neural implicit contrast-source representation: a residual MLP that takes spatial coordinates and a transmitter label as input and outputs the contrast source $\\mathbf{w}(\\mathbf{r})$ at any point. This turns an unknown discrete field into a continuous, differentiable function whose weights are the optimization variables. The argument is carried by a differentiable objective that combines the data-equation residual, the volume-integral (Lippmann–Schwinger) state-equation residual, and a total-variation term on the medium contrast $\\chi$, with automatic differentiation providing gradients through every term. Phaseless inversion fits in the same machinery because only","core_discovery":"The central claim is that choosing a continuous neural representation for the contrast source, rather than a discrete grid, materially changes the inverse-scattering problem. Parameterizing the contrast source with a residual-MLP neural field conditioned on spatial coordinates and transmitter settings makes the volume-integral-equation forward model end-to-end differentiable. The objective combines the state-equation residual, the data-equation residual, and total-variation regularization; optimizing it jointly over the MLP weights and the medium contrast is unsupervised because no ground-truth images are required. Numerical experiments are reported to show higher reconstruction accuracy and","pith_inferences":["Editorial inference: if the super-resolution claim survives independent testing, the network is effectively learning a smooth or piecewise-smooth extension of the contrast source; a Fourier-domain comparison of recovered versus true high-frequency content would clarify whether the extra detail is recovered or hallucinated.","Editorial inference: transfer to real hardware will likely hinge on calibrating the incident field and receiver responses, since the simulated measurements obey the same volume integral equation that the objective enforces; unmodeled multiple scattering could erode the reported noise-robustness gains.","Editorial inference: conditioning the neural field on transmitter settings hints that the method may extend to sparse or non-uniform antenna geometries, but the paper only demonstrates the configurations it tests."],"forward_implications":["Reconstruction cost decouples from output resolution: after optimization, the learned field can be queried on any grid, including grids finer than the one used during solving.","Phaseless measurements, where only intensities are recorded, are handled without redesigning the network or the state equation, by replacing the data misfit term.","The differentiable volume-integral objective can be combined with other differentiable priors or constraints beyond total-variation regularization.","No paired ground-truth training data are required, so applying the method to a new measurement setup means solving an optimization problem rather than retraining on a labelled dataset."],"supporting_citations":[],"fun_headline_variants":["Neural fields outperform pixel grids for inverse scattering","Unsupervised neural inversion sharpens scattering reconstruction","Continuous neural source improves full and phaseless inversion","Neural implicit source enables super-resolution scattering imaging","Differentiable neural CSI boosts accuracy and resolution"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The numerical demonstrations assume that the volume integral equation inside the objective faithfully models how the measured scattered fields were generated; if real data contain forward-model mismatch, calibration errors, or unmodeled interactions, the reported accuracy and robustness gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Neural fields outperform pixel grids for inverse scattering","Unsupervised neural inversion sharpens scattering reconstruction","Continuous neural source improves full and phaseless inversion","Neural implicit source enables super-resolution scattering imaging","Differentiable neural CSI boosts accuracy and resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1121,"prompt_tokens":725,"completion_tokens":396,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":325}},"tokens_in":469,"tokens_out":396,"duration_ms":4792,"temperature":1.0,"reasoning_tokens":325,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:22:52.262566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take calibrated experimental scattered-field data from a dielectric phantom with independently known ground-truth contrast, add controlled noise, and run the same inversion. The central claim fails if reconstruction error degrades much faster with noise than in the reported simulations, or if sampling the trained neural field on a grid finer than the training grid produces no visible resolution gain.","supporting_citations":[],"review_version":1}