{"id":"995d8dd2-b38f-4aa1-af07-9757b48ec7b6","arxiv_id":"2412.09927","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Neural field reconstruction of magnetization vector fields from probed ray transforms is more noise-robust than a discretized iterative method, with an SO(2)-equivariant variant for symmetric fields.","lead":"This paper applies smooth neural fields to vector tomography, reconstructing a 3D magnetization vector field from noisy X-ray-style projections. The method shows improved stability under noise compared with a discretized iterative baseline, especially when the field has rotational symmetry.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gradient ambiguity (Thm 1) makes full-field MSE/SSIM uninterpretable; the claimed superiority may reflect the neural net's bias on an invisible irrotational component rather than better recovery.","rationale":"The most load-bearing assumption is the implicit identification of reconstruction accuracy with full-vector-field error despite the non-injectivity of the measurement operator. This is not a style critique but a correctness issue: the loss function (12) is minimized in the data space, and Theorem 1 guarantees infinitely many fields attain the same loss. The reported MSE/SSIM then depend on the arbitrary choice of element in that equivalence class produced by the optimization. Because the neural method and the discretized baseline use different implicit regularizers (neural smoothness vs. no continuity prior), the comparison can be dominated by the unidentifiable component. The paper offers no independent support—no theorem about the neural ansatz recovering the solenoidal component, no Helmholtz-decomposition audit, no projection-domain error analysis—so the quantitative claim in the abstract is not yet grounded. The SO(2)-equivariant construction (Eq. 19) is mathematically sound and could be a useful contribution, but its demonstrated benefit is measured with the same flawed metric. The conditional verdict is therefore appropriate, and the concrete test above is needed to validate or reframe the accuracy claim.","tokens_in":7824,"tokens_out":6646,"duration_ms":75374,"concrete_test":"For each phantom, compute the Helmholtz decomposition F = F_sol + ∇φ. Reconstruct with both methods at the same noise levels. Report (i) full-field MSE, (ii) solenoidal-component MSE ||P_sol(F_rec)-F_sol||^2, (iii) projection residual 1/N Σ|Rp(F_rec)(L_i)-M_i|^2. If the neural method's advantage in (i) is not present in (ii) or (iii), then the claimed improvement over the baseline is not about reconstructable information. Also verify Table I's 'SSIM' values: since SSIM ≤ 1, values 2.38 and 2.79 indicate a plotting or calculation error that must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV evaluates reconstruction quality by comparing the reconstructed field directly to the full ground-truth vector field (Fig. 5, Table I), after explicitly acknowledging that 'the information pertaining to the exact component of the vector field is lost after the application of the ray-transform operator Rp' and that by Theorem 1 only the solenoidal component can be reconstructed. This creates an evaluation mismatch: for any field F, the single-probe data Rp(F) are identical for F and F+∇φ. Thus the error against the full F necessarily contains a positive contribution from the irrotational part that no method can remove, regardless of how well it uses the measurements. The neural model, through its smoothness prior (Eq. 13, the gradient penalty), will tend to produce reconstructions with small ∇φ; the discretized baseline has no such prior. The reported advantage in MSE/SSIM may therefore be dominated by how each method happens to behave on an unidentifiable subspace, not by fidelity of the recoverable solenoidal component. The paper never quantifies the irrotational content of its phantoms nor reports errors after projecting out the kernel (e.g., by comparing solenoidal components or by evaluating projection-domain residuals). Consequently, the central quantitative claim in the abstract—'substantially improves the accuracy'—is not supported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural-field-based method for vector tomography of magnetization vector fields, using a network F_theta to represent the field and a reconstruction loss (Eq. 12) plus a gradient-smoothness regularizer (Eq. 13). The authors argue that smooth neural representations avoid discretization artifacts and remain robust under measurement noise, and they introduce an SO(2)-equivariant architecture for fields with continuous symmetry. Numerical experiments compare the method against a discretized iterative baseline on simulated XMCD-type projections of two Bloch-point and Hopfion fields, reporting MSE and SSIM over a range of noise levels. The central quantitative claim is that the neural method substantially improves reconstruction accuracy, especially under symmetry.","tokens_in":8142,"tokens_out":3497,"duration_ms":40423,"significance":"If the claims are supported, the paper would offer a useful application of implicit neural representations to vector-field tomography, with a plausible mechanism (smoothness bias) for noise robustness and a well-motivated equivariant construction. The use of neural fields to avoid explicit voxelization and the deterministic ray-sampling strategy are sensible and align with recent NeRF-style tomographic methods. However, the experimental evidence as presented is not yet sufficient to support the central claim: the evaluation metric is mismatched to the identifiability limitation established by the paper's own Theorem 1, and the reported SSIM values are internally inconsistent. The paper also lacks the hyperparameter and implementation details needed for reproducibility. The core idea is viable, but the quantitative comparison must be redone.","major_comments":[{"comment":"The quantitative evaluation is mismatched to the identifiability result. Theorem 1 states that the single-probe ray transform Rp has kernel im ∇, so measurements do not distinguish F from F + ∇φ. Section IV nevertheless computes MSE and SSIM against the full ground-truth vector field (Fig. 5, Table I). Any reconstruction will incur an unavoidable error on the irrotational component, and the relative performance of the neural method and the discretized baseline on that unidentifiable subspace is driven by their priors: the gradient penalty in Eq. (13) explicitly suppresses ∇φ for the neural model, while the discretized baseline has no such bias. The reported superiority may therefore reflect behavior on the kernel, not improved recovery of the solenoidal component that the data actually determine. The authors should quantify the irrotational content of their phantoms and report errors after projecting both the reconstruction and ground truth onto the solenoidal subspace (for example, comparing curl-field errors) or, at minimum, report projection-domain residuals. Without this, the abstract's claim of 'substantially improves the accuracy' is not supported as stated.","section":"Section IV, Theorem 1"},{"comment":"The SSIM values 2.38 and 2.79 in Table I exceed the maximum possible value of 1 for the structural similarity index, which indicates a computational or reporting error in the Hopfion comparison. Since this table is the only quantitative evidence for the SO(2)-equivariant advantage, the result cannot be assessed until the metric is corrected and recomputed.","section":"Table I"},{"comment":"The baseline comparison is not a faithful implementation of the method in [2]: the scalar part is replaced with the ground-truth mask, so only the vector part is optimized, while no equivalent ground-truth information is given to the neural method. The reported metrics are computed on the full vector field, so it is unclear what advantage or disadvantage this hybrid baseline confers. The authors should either run the complete method from [2] on the same data, or explicitly state and justify the modified protocol and its effect on the comparison.","section":"Section IV, footnote 1"},{"comment":"The manuscript provides no network architecture details, training hyperparameters (optimizer, learning rate, number of iterations, batch size), the regularization weight α in Eq. (14), the stratified sample count M in Eq. (16), or the number of random seeds. Without these, the numerical results cannot be reproduced, and the claimed robustness under noise cannot be independently verified. This is a load-bearing omission for a methods paper.","section":"Section III-IV"},{"comment":"Figure 5 appears to show a single noise realization per σ level, with no error bars or multiple trials. The visual difference between methods at high noise could be due to a single favorable initialization of the neural network. Multiple runs with different seeds and noise realizations are needed to establish that the observed stability is a property of the method rather than a random fluctuation.","section":"Section IV, Fig. 5"}],"minor_comments":[{"comment":"In the symmetry paragraph, 'g · F = g' should read 'g · F = F' (typo).","section":"Section III-C"},{"comment":"The sentence 'It is easy to see that a neural network constructed using (19) indeed satisfies is indeed SO(2)-equivariant' has a duplicated verb and should be corrected.","section":"Section III-C"},{"comment":"The statement 'ker Rp ≃ im ∇' would benefit from a precise domain specification, since the kernel of the probed ray transform on compactly supported vector fields consists of gradients of compactly supported potentials; the current wording '≃' leaves this implicit.","section":"Section II, Theorem 1"},{"comment":"The y-axis label '100' in both subplots is ambiguous; presumably it means 10^0, but the SSIM panel must also be checked because SSIM is bounded above by 1. The axis should be labeled clearly and the plotted values verified.","section":"Section IV, Fig. 5"},{"comment":"Several references lack complete bibliographic information (e.g., [5], [6], [8], [10], [11], [12], [16] have missing volume/page or journal fields). The citation of [16] as a bare arXiv preprint should be updated if a published version exists.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's scope is a methods-focused applied ML manuscript; it may be a reasonable fit for cond-mat.dis-nn but the editor should consider whether the experimental rigor meets the journal's standards. The main unresolved issue is the evaluation mismatch with Theorem 1; if the authors can redo the metrics on the recoverable subspace and provide full training details, a resubmission could be viable. The presence of impossible SSIM values suggests that the numerical pipeline has not been carefully checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper applies neural fields to vector tomography of magnetization, which is a reasonable transfer of the NeRF toolkit to a real problem. The SO(2)-equivariant architecture for the Hopfion case is a nice touch, and the qualitative noise robustness in Fig. 2 is genuinely suggestive.\n\nBut the quantitative claims do not hold up. The authors state Theorem 1: only the solenoidal component is recoverable from these probed ray transforms. Then they evaluate full-field MSE and SSIM against the ground truth. That measures an unidentifiable component, and the neural network's gradient penalty (Eq. 13) will suppress the irrotational part that any method cannot recover. So the numerical advantage they report could just reflect the network's bias on an invisible subspace, not better estimation of what is actually identifiable. The stress-test note is correct. They need to report errors on the solenoidal component (projected onto the kernel) or at least use projection-domain residuals.\n\nSecond, Table I reports SSIM values of 2.38 and 2.79. SSIM is capped at 1. That is not a debatable choice; it is a sign of a computational bug. This alone should have been caught before submission.\n\nThird, the experimental section is under-specified. No hyperparameters (alpha, network depth/width, learning rate, sampling count), no error bars, no code release, and a single baseline that is modified in footnote 1 (the scalar part replaced by the ground-truth mask). That baseline modification might make the comparison more favorable to the proposed method, but the lack of details and the single dataset make the robustness claim unsupported.\n\nThe method itself is plausible; the idea of a smooth neural field for vector tomography is not new in structure (NEAT, NeRF), but the probed ray-transform formulation and the equivariant variant are sensible. The paper is worth engaging with—if the authors fix the evaluation, correct the SSIM, and release code and settings, it could become a useful tool for 3D nanomagnetism.\n\nFor now, I would not cite it in this form. The intended audience is the computational imaging / 3D magnetism community, and they would get value from the architectural ideas, but only after the evidence is rebuilt.\n\nRegarding peer review: send it out. The core idea is not dead, and the flaws are fixable in revision, not fatal. But the referee should hold it to a high bar on the evaluation metrics.","headline":"Strong idea, but the quantitative evidence is compromised by the gradient ambiguity and an impossible SSIM; the method warrants peer review after major fixes.","tokens_in":8635,"tokens_out":3110,"would_cite":false,"duration_ms":33704,"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":"The paper argues that representing a magnetization vector field as a smooth neural field, rather than as a discretized voxel array, gives tomographic reconstructions that remain accurate under measurement noise, and that encoding a known…","keywords":["vector tomography","magnetization","neural fields","ray transform","SO(2)-equivariance","tomographic reconstruction","noise robustness"],"falsifier":"Compute the MSE and SSIM after Helmholtz-decomposing both the reconstruction and the ground truth into solenoidal and irrotational parts on the Bloch-point phantom, and also add a large smooth gradient field to the ground truth, which leaves all single-probe projections unchanged; if the neural method's advantage disappears when only the solenoidal components are compared, the central claim of improved physical reconstruction is refuted.","tokens_in":7619,"feed_emoji":"🧲","tokens_out":6524,"duration_ms":65604,"temperature":0.7,"pith_summary":"Reconstructing a three-dimensional magnetization vector field from X-ray or electron projections is partially ill-posed: with a single probe, fields that differ by the gradient of a scalar function produce identical measurements. The paper proposes to represent the unknown field as a neural network with smooth activations and to train it against ray-integrated projections, adding a gradient penalty to suppress noise. On a simulated Bloch-point phantom, this neural-field method keeps reconstruction error low and image similarity high as Gaussian noise grows, while a voxel-discretized iterative baseline degrades quickly. For a Hopfion sample with rotational symmetry, a network built to be rotation-equivariant reconstructs more accurately than a non-equivariant network. If the approach carries to experiments, it offers a route to noise-robust three-dimensional magnetic imaging without high-resolution memory costs.","feed_headline":"Smooth neural fields beat voxel grids in noisy magnetic tomography","feed_subtitle":"A neural-field model keeps 3D magnetization reconstructions accurate as measurement noise grows.","key_machinery":"The load-bearing object is the neural field $F_\\theta$, a multilayer perceptron with smooth activations mapping $(x,y,z)$ to a vector in $\\mathbb{R}^3$, trained with the loss $L_\\alpha = L_{\\mathrm{rec}} + \\alpha L_{\\mathrm{reg}}$. Here $L_{\\mathrm{rec}}$ approximates the probed ray transform by stratified Monte Carlo sampling along each ray, and $L_{\\mathrm{reg}}$ penalizes the field's gradient at uniformly sampled points to suppress noise-induced irregularities. For symmetric samples, the architecture enforces $SO(2)$-equivariance by rotating the network output back to the local frame, so the symmetry is exact by construction rather than learned. The theory framing the problem is Theorem 1: single-probe measurements cannot distinguish fields differing by a gradient.","core_discovery":"The central claim is that modeling the field as a continuous neural function $F_\\theta$ and solving $\\min_\\theta \\frac{1}{N}\\sum_i |R_p(F_\\theta)(L_i)-R_p(F)(L_i)|^2 + \\alpha \\sum_j |\\nabla F_\\theta(x_j)|^2$ removes the artifacts that voxel discretization introduces, and that the error advantage grows with noise. The paper reports lower mean-squared error and higher structural similarity for the neural method than for the discretized baseline at all tested noise levels, and shows that stitching the field's $SO(2)$ symmetry into the architecture reduces MSE from 0.39 to 0.14 for a Hopfion. Because $\\ker R_p \\simeq \\operatorname{im}\\nabla$, only the solenoidal part is determined by the measurements; the reconstruction is therefore understood as recovering the field up to a gradient.","pith_inferences":["If the unrecoverable gradient component stays small in practice, the same smoothness prior should carry over to experimental XMCD or Lorentz TEM data, where noise and missing-wedge artifacts dominate; this could be tested by running the same loss on experimental projections.","A fairer benchmark would compare only the solenoidal parts of reconstruction and ground truth, because single-probe measurements cannot determine the irrotational part; until that is reported, the quantitative gains over the baseline may partly reflect how each method handles an invisible component.","The equivariance trick is not limited to $SO(2)$: discrete symmetries can be enforced by group averaging, and more realistic sample symmetries such as mirror planes or dihedral rotations could be encoded the same way.","The same neural-field formulation applies to any inner-product-probe vector tomography problem, including electron holography and neutron imaging, whenever the unknown field is smooth."],"forward_implications":["At every tested noise level from $\\sigma=0$ to $\\sigma=7.7$, the neural-field reconstruction achieves lower MSE and higher SSIM than the discretized baseline, with the gap growing as noise increases.","Because the network maps coordinates directly to field values, reconstructions can be evaluated at arbitrary spatial resolution without increasing memory usage, only training time.","For $SO(2)$-symmetric samples, an equivariant network gives MSE 0.14 versus 0.39 for a non-equivariant network and higher SSIM.","Smooth activations together with the gradient regularization term suppress noise-induced artifacts, so the method remains stable even when the projections are strongly corrupted."],"supporting_citations":[{"why":"It supplies the discretized voxel-optimization baseline, the Bloch-point phantom, and the XMCD projection setup that the neural method is compared against.","marker":"[2]"},{"why":"It introduces the neural radiance field representation and stratified point sampling along rays used to approximate the ray transform $R_p(F_\\theta)(L)$.","marker":"[15]"},{"why":"It defines probed ray transforms and the condition under which a set of probes yields unique reconstruction.","marker":"[25]"},{"why":"It provides Theorem 1, the kernel $\\ker R_p \\simeq \\operatorname{im}\\nabla$, which defines what is and is not recoverable from single-probe measurements.","marker":"[26]"},{"why":"It supplies the Hopfion magnetization configuration used to test $SO(2)$-equivariant reconstruction.","marker":"[30]"},{"why":"It is referenced alongside [30] as the source of the Hopfion setup with physical constraints used in the symmetry experiment.","marker":"[13]"}],"fun_headline_variants":["Neural fields sharpen noisy magnetization tomography","Continuous neural fields fix voxel artifacts in vector tomography","Magnetization imaging goes neural, beats voxel grids under noise","Symmetry-aware neural nets cut Hopfion reconstruction error by 64%","No pixelation: neural vector tomography stays accurate in noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported MSE and SSIM gains assume that the neural network does not introduce a large spurious gradient component, even though the single-probe measurements used here cannot constrain that component at all.","fun_headline_variants_meta":{"raw":{"variants":["Neural fields sharpen noisy magnetization tomography","Continuous neural fields fix voxel artifacts in vector tomography","Magnetization imaging goes neural, beats voxel grids under noise","Symmetry-aware neural nets cut Hopfion reconstruction error by 64%","No pixelation: neural vector tomography stays accurate in noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1525,"prompt_tokens":821,"completion_tokens":704,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":622}},"tokens_in":437,"tokens_out":704,"duration_ms":6638,"temperature":1.0,"reasoning_tokens":622,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:33:55.476143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the MSE and SSIM after Helmholtz-decomposing both the reconstruction and the ground truth into solenoidal and irrotational parts on the Bloch-point phantom, and also add a large smooth gradient field to the ground truth, which leaves all single-probe projections unchanged; if the neural method's advantage disappears when only the solenoidal components are compared, the central claim of improved physical reconstruction is refuted.","supporting_citations":[{"cited_title":"Tomographic reconstruction of 3-d vector fields using inner product probes,","cited_arxiv_id":null,"evidence_quote":"It defines probed ray transforms and the condition under which a set of probes yields unique reconstruction."},{"cited_title":"X-ray tomography of one- forms with partial data,","cited_arxiv_id":null,"evidence_quote":"It provides Theorem 1, the kernel $\\ker R_p \\simeq \\operatorname{im}\\nabla$, which defines what is and is not recoverable from single-probe measurements."},{"cited_title":"Three-dimensional magnetization reconstruction from electron optical phase images with physical constraints,","cited_arxiv_id":null,"evidence_quote":"It is referenced alongside [30] as the source of the Hopfion setup with physical constraints used in the symmetry experiment."}],"review_version":1}