{"id":"6d9356ea-e4c5-456d-b1fb-c3fa61f39a02","arxiv_id":"2412.19381","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single-image, physics-informed neural network can reconstruct spatially varying in-plane and 3D magnetization textures from simulated magnetic field images and can distinguish Bloch from Néel skyrmions and Bloch handedness.","lead":"Researchers trained a physics-based neural network to reconstruct the magnetic structure of thin magnetic films from images of their stray magnetic fields. The method can map magnetization directions that vary across the image and can tell apart different types of skyrmions, small swirling magnetic textures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section V's skyrmion-type and Bloch-helicity discrimination is confounded by the additive α=0.8 initial guess in Eq. (6); the reported >3σ separation may be inherited from the candidate's forward field rather than from the NN reconstruction, and no zero-NN baseline is reported.","rationale":"The reader's weakest assumption is in the right place: Sec. V and Eq. (6) are the load-bearing part of the central claim. I partially agree with the reader. My sharper concern is that the classification may be an artifact of the additive αM_init term rather than a property of the NN reconstruction; this is not a question of noise robustness alone but of what quantity is actually being measured. The other contributions (masked reconstruction, multi-component input, fidelity metrics) are plausible and internally consistent; the forward model is standard and the simulations are informative. The lack of code/data and the absence of a Bayesian comparison weaken reproducibility but do not by themselves invalidate the argument. I do not think the paper should be rejected; the right outcome is the reader's conditional acceptance, with the added condition that the zero-NN baseline and candidate-mismatch test be reported. Hence I keep the verdict unchanged.","tokens_in":16741,"tokens_out":10250,"duration_ms":103336,"concrete_test":"Run the Sec. V cluster analysis on the same four simulated skyrmion classes with (i) a zero-NN baseline: compute η_B and SSIM for M'_NN = αM_init + σ with the network output frozen at zero; (ii) perturbed candidates: diameter ±20%, center shifted by one pixel, α ∈ {0.6,0.8,1.0}, at SNR=5 and 10 with the same 20-run protocol. If the baseline separates the classes as well as the trained NN, or if the correct-type cluster drops below 3σ separation under perturbations, the Bloch-helicity claim should be weakened to conditional pending experimental validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim to test is the abstract's 'can even identify the helicity of Bloch Skyrmions.' In Sec. V the discriminator is built from Eq. (6), M'_NN = M_NN + αM_init + σ, with α=0.8 and a scaled, noisy copy of the true texture as the correct initial guess. Because 80% of the final magnetization is fixed by the candidate before optimization, the field error and the SSIM with M_init are both strongly influenced by the candidate itself. The paper reports the clusters of (η_B, SSIM) only after NN training and never reports the field error of αM_init alone (M_NN frozen at zero). If that zero-NN baseline already separates Bloch-left from Bloch-right (and from Néel guesses), then the experiment demonstrates template matching with a forward model, not an NN reconstruction capability. Additionally, the candidates are perfectly registered to the true texture (same diameter, position, amplitude), and only four classes, SNR=10, 20 runs, and 100 epochs are tested. Appendix H shows the same procedure cannot separate Néel handedness, so the helicity claim is narrow and unvalidated against realistic candidate mismatch. For the central claim to hold, the separation must survive candidate perturbations (±20% diameter, pixel shift, unknown α) and experimental unknowns; this is not established by the current evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents an untrained, physics-informed neural network method for reconstructing magnetization distributions from magnetic stray-field images, extending prior work to spatially varying magnetization directions. The method uses a convolutional network whose output is forward-propagated to compute a magnetic field that is compared with the measured field; it incorporates weighted masks, multi-component field inputs, and initial magnetization guesses. The authors demonstrate the approach on simulated examples including an 'Echidna' with orthogonal in-plane magnetization, a canted platypus crossing the image boundary, overlapping hallbars, a rotating torus, skyrmions, and merons. They further propose using initial guesses to identify skyrmion type and Bloch helicity from the reconstruction error and SSIM relative to the guess.","tokens_in":17070,"tokens_out":4095,"duration_ms":36935,"significance":"If valid, the method addresses a real limitation of Fourier-based reconstruction, which is ill-posed for in-plane magnetization and struggles with spatially varying textures. The multi-component and mask extensions are natural and well motivated, and the all-simulation setting permits quantitative comparison with ground truth. The paper's strengths are the clear forward model, the concrete demonstration on several non-trivial textures, and the explicit caveats in Appendices F-H that reconstructions of merons and Néel-handedness are imperfect. The skyrmion-type classification claim, however, rests on a diagnostic whose control baseline is missing; that claim needs additional evidence before the headline 'can even identify the helicity of Bloch Skyrmions' is established.","major_comments":[{"comment":"The initial-guess procedure of Eq. (6) adds αM_init with α=0.8 to the network output, so 80% of the final magnetization is fixed by the candidate before optimization. The clusters in Fig. 4(c,f) combine the field error and SSIM after training, but no baseline is reported with the network output frozen at zero (M_NN=0). Without that zero-NN baseline, the reported separation between Bloch-left and Bloch-right guesses may reflect the forward fields of the candidate textures themselves rather than any capability of the NN reconstruction. Please add this baseline for all four candidate classes and both Bloch handednesses; if the baseline already separates the classes, the experiment demonstrates template matching with a forward model, not NN-based reconstruction.","section":"Sec. V, Eq. (6)"},{"comment":"The classification experiment is restricted to perfectly registered candidate textures with the same diameter, position, amplitude, and standoff as the ground truth, at SNR=10, with 20 runs and 100 epochs. Appendix H shows the same procedure cannot separate Néel handedness, so the helicity claim is narrow. To support the claim that the method can identify skyrmion type and Bloch helicity in experimental images, the separation must be shown to survive realistic candidate mismatch (e.g., ±20% diameter, pixel shift, unknown α, unknown standoff, and unknown background). Without such a robustness study, the method is not demonstrated for the stated application.","section":"Sec. V, Figs. 4 and 8"},{"comment":"Reported fidelity and SSIM values (e.g., F=0.91 in §IV, SSIM values in Appendix E) are single numbers without error bars, and the 'greater than 3σ separation' in §V is not defined. Please report the mean and standard deviation over the 20 runs for each metric, and state how the σ in the 3σ claim is computed. This is needed to judge whether the classification clusters are statistically significant, especially because Fig. 8 shows overlap for the Néel case.","section":"Sec. IV and Sec. V"}],"minor_comments":[{"comment":"The abstract twice uses 'ill-poised'; the standard term is 'ill-posed'.","section":"Abstract"},{"comment":"The text states 'SSIM(Mx) = 0.986 and SSIM(Mx) = 0.969'; the second instance should refer to SSIM(My).","section":"Appendix E"},{"comment":"The symbol σ is used for the random noise in Eq. (6), for standard deviation in Appendix A, and for the spatial filter width in Appendix A; please use distinct symbols for these quantities.","section":"Eq. (6) and Appendix A"},{"comment":"The architecture table omits activation functions, padding, and upsampling layers, and the optimizer and learning rate are not stated; please complete the training details so the results are reproducible.","section":"Table I and Section II"},{"comment":"The notation 'Bxyz' is undefined; please write explicitly that it denotes the decomposed Bx, By, and Bz field components.","section":"Fig. 3 caption"},{"comment":"The quantities η_B and η_M are used in the text and figures, but only η_M is defined in Eq. (4); please define η_B as well.","section":"Sec. V and Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The reconstruction of non-uniform magnetization textures is a useful contribution and the simulations are well grounded in the forward model. The main concern is the skyrmion-type and helicity classification in Sec. V, which currently lacks a no-network baseline and robustness tests; this is fixable within the manuscript's scope by adding the requested control experiments. I do not see grounds for rejection, but the classification claim should not be published in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the multi-component extension of the untrained PINN is a real step forward and the simulations are clean enough to take seriously, but the Section V skyrmion-classification claim is not yet supported because the α=0.8 initial guess does most of the work.\n\nWhat is genuinely new: the paper removes the uniform-direction constraint that limited the earlier Dubois work by adding Mx/My (and Mz) output channels, and it shows that masks and multi-component field input meaningfully improve reconstruction. The forward model is standard magnetostatics, and the simulated benchmarks with known ground truth are appropriate for this stage. The comparison against the Fourier method is fair and the improvements are plausible. I also credit the paper for putting failed cases in the appendices — the Meron reconstruction and the Néel handedness test are not swept under the rug.\n\nThe soft spot is the helicity discrimination. Eq. (6) sets M'_NN = M_NN + αM_init + σ with α=0.8 on the true texture, so 80% of the final magnetization is fixed before the network does anything. The reported >3σ separation of Bloch-left from Bloch-right in Fig. 4 may be inherited from the forward field of the initial guess. The paper never shows the zero-NN baseline, i.e. the field error and SSIM of αM_init alone. Without that, the experiment reads as template matching with a forward model, not as evidence of an NN reconstruction capability. The test conditions are also narrow: four simulated texture classes, SNR=10, 20 runs, 100 epochs, and candidates perfectly registered to the true texture. Appendix H already shows the same procedure cannot separate Néel handedness, so the abstract's \"identify the helicity of Bloch Skyrmions\" is narrower than it sounds. None of this breaks the main reconstruction contribution, but it does break the headline demo as stated.\n\nTwo smaller things. There is no code or data release, which matters for a methods paper. And the paper cites arXiv:2411.18882 but does not compare to it; if that work already reconstructs non-uniform magnetization profiles, the novelty statement is overstated and needs an explicit discussion.\n\nWho this is for: people using NV or widefield magnetic microscopy who want quantitative magnetization maps and are willing to accept simulation-only validation for now. I would send it to review, but with a clear request for the zero-network baseline, realistic candidate mismatch tests, and either code/data or a stronger justification for withholding them.","headline":"Useful multi-component PINN extension for magnetization reconstruction, but the skyrmion-helicity discrimination is confounded by the 80% initial guess and needs a zero-network baseline before the headline claim is credible.","tokens_in":17555,"tokens_out":2538,"would_cite":true,"duration_ms":23310,"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 neural network that fits a single magnetic field image can reconstruct spatially varying magnetization, including the type and handedness of skyrmions.","keywords":["magnetization reconstruction","neural network","inverse problem","skyrmion","stray field imaging","physics-informed neural network","NV magnetometry","magnetic texture"],"falsifier":"Take an experimental stray-field image of a skyrmion whose Bloch/Néel type and handedness are already known from an independent technique, run the reconstruction with all four initial guesses, and check whether the correct guess forms a cluster separated by more than $3\\sigma$ in field error and SSIM; the central claim fails if the correct guess does not win. A cheaper check is the paper's own simulated test at SNR below 10 or with a standoff error, where the four-type clusters may overlap.","tokens_in":1521,"feed_emoji":"🧲","tokens_out":6520,"duration_ms":91058,"temperature":0.7,"pith_summary":"Converting a magnetic field image into a magnetization map is an ill-posed inverse problem, and the standard Fourier-space inversion becomes unreliable when magnetization varies in direction inside the image. This paper argues that an untrained, physics-informed neural network can perform that inversion by fitting a single image: the network outputs a magnetization guess, Maxwell's equations forward-propagate it to a field, and the difference from the measured field is minimized. Several features make this work for textures: a mask that restricts where magnetization can appear, a loss mask that focuses the fit, feeding all three vector field components instead of one, and an initial guess that steers the network toward the correct spin texture. On simulated skyrmions, the approach separates Bloch from Néel textures and identifies Bloch handedness; it also reconstructs a rotating magnetization pattern and a meron with moderate fidelity. If this holds, complex spin textures in 2D magnets could be characterized quantitatively from a single stray-field image without training data.","feed_headline":"Neural nets turn one magnetic field image into full magnetization map","feed_subtitle":"Untrained physics-informed network with masks reconstructs skyrmions and mixed in-plane textures.","key_machinery":"The load-bearing object is the forward map $B = A M$, the Fourier-space expression of Maxwell's equations for a quasi-2D magnetization layer, which is non-invertible and therefore needs an approximate inverse. The paper's network is an untrained U-net that outputs magnetization components $M_x$, $M_y$, and for 3D textures $M_z$; those are forward-propagated to a field, and an error function compares the result to the measured image. Three additions carry the argument: a spatial mask $\\zeta_S$ applied to the output (Eq. 2), a loss mask $\\zeta_L$ weighting where the error is computed (Eq. 3), and an initial-guess modification $M'_{NN} = M_{NN} + \\alpha M_{init} + \\sigma$ (Eq. 6) that pins the optimization near a candidate spin texture. The multi-component input exploits the fact that a single measured field component can be Fourier-decomposed into all three vector components, which restricts the solution space and suppresses the cross-talk between $M_x$ and $M_y$.","core_discovery":"This paper's central claim is that neural network reconstruction can recover the magnetization of a sample whose magnetization direction varies from pixel to pixel within a single image, a task the paper says was not possible with previous Fourier-based or NN methods. The network is never trained on examples; it learns by fitting one image, using the forward relation between magnetization and stray field as the physical constraint. With weighted source masks, multi-component ($B_x$, $B_y$, $B_z$) field input, and a noisy, scaled copy of the expected spin texture as an initial guess, the reconstruction reaches fidelity $F = 0.91$ on overlapping in-plane structures and can differentiate Bloch from Néel skyrmions, including left- versus right-handed Bloch textures.","pith_inferences":["The initial-guess procedure is effectively a model-selection scheme: each candidate spin texture defines a preferred solution neighborhood, and the network is asked which neighborhood fits the data best; the same logic could be applied to chiral domain walls or vortex core polarities.","Because the paper finds that Néel handedness cannot be separated while Bloch handedness can, the stray-field signature of Néel chirality at the tested standoff and SNR is likely too weak for this inverse method; a lower standoff or higher SNR might recover it, but the paper does not demonstrate that.","The method's reliance on a good initial guess means that for a completely unknown texture, a multi-start search over many candidate guesses would be needed; the paper uses four candidates.","Since the same forward map governs current-density imaging, the mask and multi-component improvements should transfer to reconstructing current distributions from magnetic images."],"forward_implications":["A single magnetic field image can yield quantitative, spatially resolved magnetization maps for samples with in-plane, out-of-plane, or mixed magnetization directions, without any training data.","Bloch and Néel skyrmions can be distinguished, and left- versus right-handed Bloch skyrmions identified, by running the reconstruction with each candidate texture as an initial guess and comparing field error and SSIM.","Source masks suppress the background-magnetization artifacts that plague Fourier inversion, especially for partially imaged flakes.","Using all three decomposed magnetic field components increases reconstruction fidelity from 0.57 for a single component to 0.91 for multi-component input in the overlapping-sample test.","The method transfers across sensor types, since a single-sensor image is decomposed to vector components before fitting."],"supporting_citations":[{"why":"Supplies the untrained, physically informed neural network approach that this work extends from uniform magnetization to spatially varying magnetization.","marker":"[42]"},{"why":"Provides the Fourier-space forward model B = AM and the decomposition of a single measured field component into vector field components.","marker":"[19–21]"},{"why":"Documents the noise amplification and artifacts of Fourier inversion for in-plane magnetization, motivating the neural network alternative.","marker":"[25]"},{"why":"Shows how a single magnetic field image is decomposed into Bx, By, and Bz for multi-component input.","marker":"[24]"},{"why":"Introduces the reflective boundary method whose limitations for in-plane magnetization motivate explicit physical masks.","marker":"[48]"},{"why":"Defines the SSIM metric used to compare initial guesses and reconstructions in the skyrmion-identification procedure.","marker":"[54]"}],"fun_headline_variants":["Neural net turns a magnetic field image into a full magnetization map","Single-image neural fit reconstructs mixed in-plane and skyrmion textures","Untrained, physics-informed network solves ill-posed magnetization inversion","From stray field to spin texture: one neural fit does it all"],"cache_read_input_tokens":19712,"weakest_assumption_plain":"The skyrmion-identification claim rests on the assumption that starting from a slightly smeared, noisy copy of the true spin texture reliably guides the network to the true magnetization, and that wrong starting guesses lead to measurably worse fits even in real experimental noise.","fun_headline_variants_meta":{"raw":{"variants":["Neural net turns a magnetic field image into a full magnetization map","Single-image neural fit reconstructs mixed in-plane and skyrmion textures","Untrained, physics-informed network solves ill-posed magnetization inversion","From stray field to spin texture: one neural fit does it all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00055,"raw_usage":{"total_tokens":2558,"prompt_tokens":810,"completion_tokens":1748,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":1673}},"tokens_in":426,"tokens_out":1748,"duration_ms":11964,"temperature":1.0,"reasoning_tokens":1673,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:38:56.087694+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an experimental stray-field image of a skyrmion whose Bloch/Néel type and handedness are already known from an independent technique, run the reconstruction with all four initial guesses, and check whether the correct guess forms a cluster separated by more than $3\\sigma$ in field error and SSIM; the central claim fails if the correct guess does not win. A cheaper check is the paper's own simulated test at SNR below 10 or with a standoff error, where the four-type clusters may overlap.","supporting_citations":[{"cited_title":"Prato and L","cited_arxiv_id":null,"evidence_quote":"Supplies the untrained, physically informed neural network approach that this work extends from uniform magnetization to spatially varying magnetization."},{"cited_title":"Casola, T","cited_arxiv_id":null,"evidence_quote":"Documents the noise amplification and artifacts of Fourier inversion for in-plane magnetization, motivating the neural network alternative."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows how a single magnetic field image is decomposed into Bx, By, and Bz for multi-component input."},{"cited_title":"Machine Learning for Improved Current Density Reconstruction from 2D Vector Magnetic Images","cited_arxiv_id":"2407.14553","evidence_quote":"Introduces the reflective boundary method whose limitations for in-plane magnetization motivate explicit physical masks."}],"review_version":1}