{"id":"ce598ba6-e4a0-4b20-8ca5-24743cd31d94","arxiv_id":"2608.09382","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CRPDNN uses a coordinate-residual convolutional network optimized by data consistency to reconstruct 3D contrast distributions, beating CSI and L2/3-FBE-WCIE on reported synthetic and Fresnel cases.","lead":"A new physics-driven neural network, CRPDNN, reconstructs 3D electromagnetic images directly from scattered field data without needing a preliminary guess. On synthetic and Fresnel test cases, it reports lower reconstruction error and shorter runtime than two classical inversion solvers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 5.5x and 12.1x speedups over CSI and L2/3-FBE-WCIE are not yet controlled comparisons, because the paper does not disclose whether the baselines ran on the same CPU/GPU and with matched stopping criteria.","rationale":"I identify the runtime-comparison fairness as the single most load-bearing concern because the strongest claim includes speedups as a headline result, and the paper provides no evidence that the baselines were run under identical conditions. The reader's weakest assumption (optimization convergence) is also important, but it is partially mitigated by the empirical results: low relative errors across all seven 3-D cases suggest the optimizer did not stall for these examples, and the claim is explicitly scoped to 'reported cases.' The hardware/implementation ambiguity, by contrast, is a controlled-experiment issue that can invalidate the speed claim without any error in the optimization. I therefore disagree with the reader's choice of weakest assumption, though the paper's overall conditionality remains appropriate. The concrete test above would settle whether the speedups are algorithmic or hardware-driven.","tokens_in":11549,"tokens_out":8151,"duration_ms":79613,"concrete_test":"Check the linked repository for the exact CSI and L2/3-FBE-WCIE implementations and any runtime scripts, then re-run all three methods on the same workstation under three conditions: (i) all on CPU, (ii) CRPDNN on GPU and baselines on CPU (as likely originally), (iii) all on GPU if the baselines can be GPU-accelerated, with matched stopping criteria (same relative residual tolerance and maximum iterations). Record wall-clock time and per-iteration cost for each method. If condition (i) reduces the speedups to below 2x, the headline claim should be revised to specify the hardware dependence; if condition (iii) is infeasible, the paper should state the CPU-only baseline times and explicitly label the comparison as CPU-vs-GPU.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim's speed component rests on wall-clock runtimes (~133 s for CRPDNN vs ~729 s and ~1614 s for the baselines), but the paper never states which device each method used. CRPDNN is a 48-channel 3-D convolutional network trained with Adam on an RTX 4090, whereas CSI and L2/3-FBE-WCIE are classical FFT-accelerated iterative solvers that are typically implemented on CPU. If the baselines were CPU-only, the comparison conflates algorithmic efficiency with hardware acceleration. Additionally, the paper does not report baseline iteration counts, stopping tolerances, or whether the same codebase was used for all methods, and the GitHub link (github.com/Physics-driven-methods) is a group account rather than a specific repository. A speedup claim this central needs a controlled benchmark on identical hardware with matched convergence criteria. The accuracy improvement (2.10% vs 7.97% vs 3.99%) is less affected by hardware, but the speed claim is independently load-bearing in the abstract and conclusion. The reader's convergence concern is valid but secondary: all seven reported synthetic cases show low errors, so a local-minimum failure would be visible in the numbers; the speed comparison, by contrast, is unverifiable from the paper alone.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-driven neural network solver, CRPDNN, for 3-D electromagnetic inverse scattering. The network is initialized with zero contrast plus a normalized-coordinate encoding, parameterizes the unknown complex contrast distribution with a dilated residual convolutional architecture, and is optimized per measurement by minimizing a data-consistency loss together with adaptive bound and total-variation regularization. No labeled training data or preliminary reconstruction is used. The authors report quantitative comparisons with classical iterative solvers (CSI and L2/3-FBE-WCIE) on 3-D synthetic cases, including noise-contaminated measurements, a supplementary 2-D comparison with existing PDNN baselines, and reconstructions from the 3-D Fresnel experimental database. The central claims are that CRPDNN achieves lower relative reconstruction error and substantially shorter runtime than the classical baselines, and that it avoids the instability of subregion-selection-based PDNN frameworks.","tokens_in":11789,"tokens_out":2757,"duration_ms":32192,"significance":"If the claims hold, the paper makes a useful contribution: it shows that a per-sample physics-driven network without preliminary reconstruction, labeled data, or explicit subregion selection can solve nonlinear 3-D inverse scattering problems with accuracy competitive with or better than classical iterative methods. The forward model and loss formulation are standard and clearly stated, the loss weights and architecture constants are fixed across experiments, and the quantitative tables support the accuracy claim for the reported cases. The 2-D comparison with PDNN-Full, PDNN-Subregion, IPDNN-InitFull, and IPDNN is informative and gives concrete evidence on the failure modes of subregion selection. However, the runtime advantage is not yet a controlled comparison, and the noise-robustness experiments lack statistical reproducibility because only one noise realization per SNR is reported.","major_comments":[{"comment":"The central speedup claim (roughly 5.5-fold and 12.1-fold over CSI and L2/3-FBE-WCIE) is not supported by a controlled benchmark. The paper states that all methods were evaluated on a workstation with an Intel i9 CPU and an RTX 4090 GPU, but it does not state which device each method used, nor the iteration counts, stopping tolerances, or convergence criteria for CSI and L2/3-FBE-WCIE. Since CRPDNN is a GPU-oriented convolutional network while the classical baselines are typically CPU-oriented FFT-accelerated solvers, the reported wall-clock comparisons may conflate algorithmic efficiency with hardware acceleration. The authors should rerun or report baselines on the same device as CRPDNN (or both devices), with matched stopping criteria, and disclose the exact code versions. Without this, the speed component of the abstract's central claim is not verifiable.","section":"Section 3.4; Tables 1-4"},{"comment":"The noise-robustness experiments report only one noise realization per SNR (20 dB and 10 dB). Because additive white Gaussian noise is random, the relative errors in Tables 3 and 4 may vary substantially across realizations, especially for the smaller structures in Cases 2 and 3. The authors should report means and standard deviations (or median and interquartile range) over at least several independent noise realizations, and state whether the reported runtimes correspond to a single run or an average. This is needed to support the claim that CRPDNN 'maintains reliable reconstruction performance under noisy measurements.'","section":"Section 4.3; Tables 3-4"},{"comment":"The method minimizes a nonconvex data-consistency loss over the network parameters using Adam for a fixed 300 iterations from a zero-contrast initialization, but the paper provides no convergence analysis, loss-curve reporting, or failure-mode study. All seven reported 3-D synthetic cases converge to low-error reconstructions, which is encouraging, but the absence of any failure cases makes it hard to assess robustness of the 'no preliminary reconstruction' design. I ask the authors to report final data-consistency loss values and convergence curves for each case, and to test sensitivity to the random seed of Adam, the learning-rate schedule, and the 300-iteration budget. This would directly address whether the fixed optimization budget is load-bearing for the reported accuracy.","section":"Sections 3.3-3.4"},{"comment":"The Fresnel experimental results report relative reconstruction errors of 2.54% and 0.59%, but the paper does not explain how the ground-truth contrast for these measured targets is defined on the 32x32x32 voxel grid, how the forward model is calibrated to the experimental setup (including the plane-wave approximation and receiver positions), or whether these error values are comparable to the synthetic-case errors. Since the experimental section is used to claim practical applicability, the metric definition and any calibration steps should be described.","section":"Section 5"}],"minor_comments":[{"comment":"The phrase 'A WGN' should be 'AWGN'; similarly, 'L2/3-FBE-WCIE' appears with inconsistent spacing and subscript styling.","section":"Abstract and Section 1"},{"comment":"The code availability statement points to a group GitHub account (github.com/Physics-driven-methods) rather than a specific repository, which makes the claimed code release difficult to verify. A direct repository URL or DOI would be preferable.","section":"Section 3.4"},{"comment":"The table header contains formatting artifacts such as 'PDNN-F ull' and 'IPDNN-InitF ull'; the column headers should be cleaned.","section":"Table 5"},{"comment":"The comparison with PDNN frameworks uses 2-D examples from a different configuration than the 3-D experiments; the authors state this is for fair comparison, but they should explicitly caution that the 2-D CRPDNN runtimes (8 s) are not directly comparable to the 3-D runtimes reported earlier.","section":"Section 4.4"},{"comment":"The visualizations in Figures 3-6 use opacity thresholds that may obscure weak scatterers; reporting the exact thresholding procedure in the text is helpful, but the figures would benefit from also showing a fixed slice without opacity-based suppression.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's accuracy claims appear sound for the reported cases, and the method is a reasonable contribution to physics-driven inverse scattering. The main concern is the uncontrolled runtime comparison and the lack of statistical detail in the noise experiments. I would be willing to accept after a revision that addresses these points with concrete benchmark details and error-bar reporting. The citation pattern is notably self-referential among the authors' prior PDNN papers, but the related work discussion is adequate; no further action needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is real: representing the unknown contrast as a coordinate-residual function and optimizing a dilated residual CNN against the data-consistency loss, with no preliminary reconstruction, is not present in the cited PDNN papers. And it appears to work on the reported cases. The accuracy gains over CSI and L2/3-FBE-WCIE are consistent across seven synthetic 3D cases and two Fresnel targets, and the Fresnel results (2.54% and 0.59% error) give the paper a genuine proof of concept.\n\nWhat it does well: the forward model is standard FFT-accelerated VIE, the loss terms are clearly stated, and the scale-balanced adaptive weighting is a sensible way to avoid hand-tuning. The 2D comparison against four PDNN variants is the most informative section: CRPDNN avoids the catastrophic subregion-selection failures (432% and 376% errors) while staying competitive on the easy cases. The code link is to a group account, not a repo, which is a minor but real gap.\n\nSoft spots, in order of importance. First, the speedup claim (5.5x and 12.1x) is not controlled. The paper says all methods ran on the same workstation, but never states whether CSI and L2/3-FBE-WCIE used the CPU or GPU. CRPDNN is a 48-channel 3D CNN on an RTX 4090; the classical solvers are FFT-accelerated and almost certainly CPU-only. That conflates algorithmic efficiency with hardware acceleration. The runtimes need to be re-measured on identical hardware with matched stopping criteria before the abstract numbers are taken at face value. Second, baseline hyperparameters are missing: no iteration counts, tolerances, or whether the baselines ran to convergence. Third, no error bars and single noise realizations. Fourth, the convergence concern about local minima is real but secondary, because all reported cases show low errors; a trivial solution would show up in the numbers.\n\nWho this is for: computational electromagnetics and inverse-scattering researchers who want a fast single-case solver without labeled training data. The paper deserves a serious referee. I would accept it for review, ask for a controlled benchmark, error bars, and a specific code link. The central idea is sound, and the evidence, once cleaned up, should back the claims.","headline":"A genuinely new PDNN twist that likely works on the reported cases, but the speed-up numbers need a controlled hardware-matched benchmark before they can be trusted.","tokens_in":12347,"tokens_out":2046,"would_cite":true,"duration_ms":21756,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-driven network reconstructs 3-D permittivity directly from scattered fields, without a preliminary image, at 2.10% average error versus 7.97% for CSI and 3.99% for L2/3-FBE-WCIE, with 5.5- and 12.1-fold speedups.","keywords":["electromagnetic inverse scattering","physics-driven neural network","3-D volumetric imaging","coordinate encoding","residual convolutional network","contrast source inversion","total variation regularization","complex permittivity reconstruction"],"falsifier":"A reader could rerun CRPDNN on the 10 dB noisy cases for 600 iterations with several random network initializations and compare the spread of final reconstructions and data-consistency values; if the spread is comparable to the reported error gaps between CRPDNN and the baselines, or if doubling the iteration budget changes the reported errors materially, then the fixed 300-iteration schedule and initialization are load-bearing rather than the network representation itself.","tokens_in":11324,"feed_emoji":"📡","tokens_out":10829,"duration_ms":88308,"temperature":0.7,"pith_summary":"The paper proposes CRPDNN, a physics-driven neural network that reconstructs the unknown permittivity contrast of a 3-D object directly from measured scattered fields. Unlike earlier accelerated PDNN frameworks, it does not first compute a rough reconstruction to choose a subregion; instead the whole volume is represented by a residual convolutional network whose input includes normalized spatial coordinates, and only the network weights are optimized for each measurement by minimizing a data-consistency loss. On the reported noise-free 3-D synthetic cases, CRPDNN attains an average relative error of 2.10%, compared with 7.97% for CSI and 3.99% for L2/3-FBE-WCIE, while being roughly 5.5 and 12.1 times faster. The paper also shows the method keeps the lowest error among the three solvers under 20 dB and 10 dB noise and reconstructs experimental 3-D targets from a public database. If correct, this demonstrates that a physics-driven network can outperform classical iterative solvers in both accuracy and speed for volumetric electromagnetic imaging without labeled training data.","feed_headline":"3-D scattering imaged at 2.1% error, 12x faster","feed_subtitle":"Physics-driven network skips the rough first guess and still beats iterative solvers on accuracy and runtime.","key_machinery":"The carrying object is the coordinate-residual network: an 18-channel input formed by concatenating two zero-valued contrast channels with a 16-channel coordinate encoding $\\gamma(\\mathbf{r}) = \\{x, y, z, r^2, \\sin(n\\pi\\xi), \\cos(n\\pi\\xi)\\}_{\\xi \\in \\{x,y,z\\}, n=1,2}$, processed by ten dilated residual blocks with dilation rates $[1,1,2,2,4,4,8,4,2,1]$ and a residual scaling factor of $0.1$, and mapped to a two-channel output giving the real and imaginary parts of the complex contrast. The coordinate encoding provides explicit spatial information, so the network can represent the volumetric contrast without any preliminary reconstruction; the dilated blocks enlarge the receptive field while preserving voxel resolution. The optimization loss is the normalized data-consistency term $\\|\\mathbf{E}^{\\mathrm{sca}}_{\\mathrm{mea}} - \\hat{\\mathbf{E}}^{\\mathrm{sca}}(\\hat{\\chi})\\|^2 / \\|\\mathbf{E}^{\\mathrm{sca}}_{\\mathrm{mea}}\\|^2$ plus a ReLU lower-bound penalty on the real part of the contrast and a total-variation term on its magnitude, with the two regularization weights adapted per iteration by scale balancing.","core_discovery":"The central claim is that the unknown complex contrast in a 3-D electromagnetic inverse scattering problem can be represented as the output of a residual convolutional network fed with normalized spatial coordinates and zero initial contrast, and that optimizing this network's weights by a data-consistency loss yields reconstructions more accurate and much faster than classical iterative solvers. In the reported tests the average relative error is 2.10% versus 7.97% for CSI and 3.99% for L2/3-FBE-WCIE, with about 5.5- and 12.1-fold speedups; the advantage persists under 20 dB and 10 dB noise and on measured 3-D experimental data, where the two targets are reconstructed at 2.54% and 0.59% relative error.","pith_inferences":["An implicit consequence the authors do not spell out: because CRPDNN optimizes per measurement rather than training on a population, its practical value hinges on whether the fixed optimization budget and network width scale to larger DOI discretizations and higher frequencies without re-tuning.","A testable extension: apply CRPDNN to recover both permittivity and conductivity, or to multi-frequency data, by augmenting the coordinate encoding with frequency channels; the paper's experiments only cover co-polarized data.","If the reported speedups hold on larger problems, the coordinate-residual representation could be combined with progressive voxel refinement or multigrid-style continuation to push toward real-time 3-D imaging, a direction the paper mentions as future work but does not demonstrate.","The adaptive scale-balanced weighting implies that per-sample tuning of regularization strength should be unnecessary; a reader could probe this by testing targets whose contrast magnitude lies far outside the tested range."],"forward_implications":["On the reported noise-free 3-D synthetic suite, CRPDNN reaches 2.10% average relative error versus 7.97% for CSI and 3.99% for L2/3-FBE-WCIE, with about 5.5- and 12.1-fold runtime reductions.","CRPDNN reconstructs complex-valued contrast, including both real and imaginary parts of relative permittivity, and under AWGN at 20 dB and 10 dB it keeps the lowest relative error among the three solvers.","Because no preliminary reconstruction or subregion selection is needed, CRPDNN avoids the failure mode that produced relative errors of 375-432% in subregion-based PDNN baselines on difficult 2-D cases.","On measured 3-D data, the method reconstructs the two tested targets at 2.54% and 0.59% relative error, indicating operation on real experimental measurements rather than only synthetic cases.","The fixed 300-iteration optimization makes runtime roughly constant per case independent of target complexity, about 8 s in the 2-D comparison and 72-222 s in the 3-D synthetic cases."],"supporting_citations":[{"why":"Provides the contrast source inversion baseline whose error and runtime are compared.","marker":"[12]"},{"why":"Supplies the L2/3-FBE-WCIE baseline and its regularization scheme used as the second classical comparison.","marker":"[32]"},{"why":"Introduces the physics-driven neural network framework that CRPDNN extends and compares against.","marker":"[26]"},{"why":"Defines the improved physics-driven network and subregion-identification baseline used in the 2-D tests.","marker":"[27]"},{"why":"Supplies the public 3-D measured scattering data used for the experimental validation targets.","marker":"[33]"},{"why":"Documents the measurement setup and precision of the 3-D experimental scattering data.","marker":"[34]"},{"why":"Gives the FFT-accelerated integral-equation forward solver used to compute predicted scattered fields in the data-consistency loss.","marker":"[31]"}],"fun_headline_variants":["3D EM imaging: 2.1% error, 12x faster, no initial guess","Physics-driven network skips rough start for 3D EM imaging","CRPDNN: 2.1% error, 12x speedup in 3D EM inverse scattering","No initial guess needed: 3D EM solver reaches 2.1% error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that 300 iterations of an adaptive-moment optimizer, started from zero contrast, drive the CRPDNN data-consistency loss to a physically correct contrast rather than to a local minimum or a trivial solution, a step the paper does not analyze.","fun_headline_variants_meta":{"raw":{"variants":["3D EM imaging: 2.1% error, 12x faster, no initial guess","Physics-driven network skips rough start for 3D EM imaging","CRPDNN: 2.1% error, 12x speedup in 3D EM inverse scattering","No initial guess needed: 3D EM solver reaches 2.1% error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001022,"raw_usage":{"total_tokens":4315,"prompt_tokens":953,"completion_tokens":3362,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":3266}},"tokens_in":569,"tokens_out":3362,"duration_ms":24905,"temperature":1.0,"reasoning_tokens":3266,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:20:57.637378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could rerun CRPDNN on the 10 dB noisy cases for 600 iterations with several random network initializations and compare the spread of final reconstructions and data-consistency values; if the spread is comparable to the reported error gaps between CRPDNN and the baselines, or if doubling the iteration budget changes the reported errors materially, then the fixed 300-iteration schedule and initialization are load-bearing rather than the network representation itself.","supporting_citations":[{"cited_title":"van den Berg and Ralph E","cited_arxiv_id":null,"evidence_quote":"Provides the contrast source inversion baseline whose error and runtime are compared."},{"cited_title":"Three-dimensional electromagnetic inverse scattering imaging via FBE-CIE model withL 2/3 regularization and alternating direction method of multipliers.Appl","cited_arxiv_id":null,"evidence_quote":"Supplies the L2/3-FBE-WCIE baseline and its regularization scheme used as the second classical comparison."},{"cited_title":"Physics-driven neural network for solving electromagnetic inverse scattering problems.IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Introduces the physics-driven neural network framework that CRPDNN extends and compares against."},{"cited_title":"Im- proved physics-driven neural network for solving inverse scattering problems.IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Defines the improved physics-driven network and subregion-identification baseline used in the 2-D tests."},{"cited_title":"Free space experimental scattering database continuation: experimental set-up and measurement precision.Inverse Probl., 21(6):S117– S130, 2005","cited_arxiv_id":null,"evidence_quote":"Supplies the public 3-D measured scattering data used for the experimental validation targets."},{"cited_title":"Continuing with the Fresnel database: experimental setup and improvements in 3D scattering measurements.Inverse Probl., 25(2):024001, 2009","cited_arxiv_id":null,"evidence_quote":"Documents the measurement setup and precision of the 3-D experimental scattering data."},{"cited_title":"Contraction integral equation for three-dimensional electromagnetic inverse scattering problems.J","cited_arxiv_id":null,"evidence_quote":"Gives the FFT-accelerated integral-equation forward solver used to compute predicted scattered fields in the data-consistency loss."}],"review_version":1}