{"id":"0309b318-34cb-4193-b2f7-b1157b58b4b0","arxiv_id":"2607.07685","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"CNN and hand-crafted feature networks recover magnetic field (~3.8 µT), temperature (~0.12 K), and hysteresis branch from one quantitative magneto-optical domain map of Bi:YIG.","lead":"A single magneto-optical image of maze domains in a Bi:YIG film can be decoded by neural nets to recover applied field, temperature, and magnetic history at high precision. This shows domain patterns store a dense, multiparametric record of external conditions and history.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified to the central experimental claim as stated.","rationale":"The central experimental result is a clean, well-controlled demonstration that quantitative maze-domain maps of this particular high-quality Bi:YIG film contain multiparametric information recoverable by both black-box and interpretable models. Residuals are quantified, history is treated as a continuous regression target then discretized, and the dominant hand-crafted features (mean rotation, contrast, period, endpoint counts) are given clear physical meaning. The only material limitation is ordinary for a first demonstration—single sample, data-on-request, no external multi-sample test—and is already acknowledged by the authors. The reader’s weakest-assumption diagnosis therefore correctly identifies the boundary of the claim without identifying an internal flaw that would require a verdict change. Hence the CONDITIONAL verdict and MODERATE confidence remain appropriate; no stronger objection is warranted.","tokens_in":16816,"tokens_out":495,"duration_ms":6400,"concrete_test":"If the raw maps become available, retrain both models on a random 50 % subset of the 11 372 maps and evaluate residuals on the complementary held-out half; if µ0∆H_CNN and ∆T_CNN remain within ~20 % of the published 3.8 µT / 0.12 K figures and history classification stays perfect, the information-content claim is independently corroborated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper’s strongest claim is a controlled demonstration on one commercial epitaxial Bi:YIG MOIF under quasi-static, low-defect laboratory conditions: a single quantitative polarization map encodes enough information for simultaneous recovery of field, temperature, and history branch by both CNN and a 20-feature MLP, with the reported residuals and perfect history classification on held-out maps from the same campaign. That claim is internally supported by the described acquisition protocol, the dual inference routes, residual plots (Figs. 3–4), and the physical interpretation of the hand-crafted descriptors. The reader’s weakest assumption correctly flags the lack of multi-material or multi-sample transfer tests, but the manuscript itself already labels generalization as future work (Discussion/Conclusion) and does not assert that the learned mappings transfer without re-training. Within the stated scope there is therefore no load-bearing internal inconsistency or unsupported leap that would overturn the demonstration.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript demonstrates that a single quantitative magneto-optical polarization map of maze domains in an epitaxial Bi:YIG film with perpendicular anisotropy encodes sufficient information to reconstruct applied out-of-plane field, temperature, and magnetic-history branch. Using a Stokes-polarization camera, the authors acquire 11 372 high-contrast maps over µ0H ∈ [−1, +1] mT and T ∈ [21, 72] °C on both branches of the hysteresis loop. Two complementary inference routes are trained: a 20-dimensional hand-crafted feature vector (mean and higher moments of βM, Fourier domain period, skeletonized endpoint/branch counts, etc.) fed to a multilayer perceptron, and a seven-layer CNN operating on tiled, augmented maps. On a held-out 10 % validation set both models achieve near-zero systematic bias, with CNN residuals µ0∆H ≈ 3.8 µT and ∆T ≈ 0.12 K and perfect history classification; the feature-vector model is only modestly worse. Learning curves versus training-set size and physical interpretation of the most informative descriptors (contrast, period, endpoint density) are provided.","tokens_in":17048,"tokens_out":945,"duration_ms":12094,"significance":"If the reported residuals hold, the work supplies a carefully controlled experimental demonstration that magnetic domain morphology functions as a high-fidelity, multiparametric record of field, temperature and history. The dual CNN / physically interpretable feature-vector approach is a clear strength: it both quantifies the information content and links the decoded parameters to concrete material-dependent quantities (Ms(T), domain-wall energy balance, nucleation topology). The large, quantitative experimental dataset, residual plots with bias and RMS, and learning curves versus dataset size meet a high standard of reproducibility for the stated system. Within the scope of quasi-static, low-defect Bi:YIG the result is solid and opens a concrete route for single-image multiparametric magneto-optical sensing and data-driven studies of domain formation.","major_comments":[{"comment":"Discussion / Conclusion: the central claim is demonstrated on a single commercial epitaxial Bi:YIG MOIF under quasi-static laboratory conditions. While the manuscript correctly labels multi-material and dynamic generalization as future work, the abstract and final paragraph still frame the result as establishing magnetic texture as a general high-fidelity record. A short, explicit statement of the single-sample / single-material scope (and the consequent need for re-training) should be added so that the claim strength matches the evidence.","section":null},{"comment":"§2.2 and Supporting Information (feature definitions): morphological descriptors (N±,ends, N±,branches, fractal dimension, etc.) rely on binarization and skeletonization whose thresholds are not stated. Because these features are used both for inference and for the physical interpretation of history encoding, the precise thresholding procedure (or a sensitivity analysis) must be reported so that the endpoint-density argument can be reproduced.","section":null}],"minor_comments":[{"comment":"Figure 3 caption and text: residual units are given as µ0∆H and ∆T; a brief note that the shaded bands are RMS (not standard error of the mean) would avoid ambiguity.","section":null},{"comment":"Methods (Computing): training times and hardware are useful, but the exact tile size used for the CNN and the random-seed / split protocol for the 90/10 partition are not stated; adding them would improve reproducibility.","section":null},{"comment":"Supporting Information feature list: several figures are labeled “Fig. S4” twice (aspect-ratio block); renumber for clarity.","section":null},{"comment":"Introduction: the comparison “more than one order of magnitude” improvement over prior ML domain studies is asserted without a tabulated relative-error baseline; a short table or explicit citation of the numbers used would strengthen the claim.","section":null},{"comment":"Abstract / keywords: “magnetic history” is central yet not listed among the keywords; consider adding it.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The work is a high-quality experimental demonstration on one well-chosen system. Fit for a materials/condensed-matter journal is good; the main risk is over-generalization language that can be fixed with minor textual tightening. No integrity or novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a solid, carefully executed demonstration: a single quantitative Stokes-camera polarization map of maze domains in epitaxial Bi:YIG carries enough information to recover applied field (~3.8 µT RMS with the CNN), temperature (~0.12 K), and the hysteresis branch simultaneously, with near-zero bias and perfect history classification on held-out maps from the same campaign. That is new relative to the prior ML-on-domains literature, which mostly stayed in simulation, binarized images, single-target inference, or composition classification without joint field+temperature+history recovery from experimental quantitative maps.\n\nWhat they do well is the dual route. The CNN shows the upper bound from full spatial information; the 20-feature MLP (mean/std of βM, Fourier period, skeleton endpoints/branches, etc.) recovers most of that performance and attaches physical meaning—endpoint density tracks history, period and contrast track temperature, mean tracks magnetization. The acquisition is serious: 11k maps, controlled field/temperature, train/val split, residual plots vs. true parameters, learning curves vs. dataset size. Circularity is low; labels are independent control variables. Citations cover the relevant domain-physics and ML-on-texture work without obvious gaps.\n\nSoft spots are ordinary for a first demonstration, not load-bearing. Everything is one commercial MOIF under quasi-static, low-defect lab conditions; generalization across materials, defects, dynamics, or imaging modalities is left as future work and is not claimed. Data are “on request,” so independent re-analysis is not possible from the public record. Hyperparameters (learning rates, tile size, skeleton thresholds) are free but not hidden. None of this overturns the central claim as stated.\n\nThis is for people who care about magneto-optical sensing, domain physics, or inverse inference from modulated phases. It deserves a serious referee. I would engage with it and expect it to be useful for anyone building multiparametric MOIF readouts or thinking about information density in maze textures.","headline":"Clean experimental demo that one quantitative Bi:YIG maze map encodes field, temperature, and hysteresis branch, recovered by both CNN and interpretable features with quantified residuals.","tokens_in":17629,"tokens_out":494,"would_cite":true,"duration_ms":5775,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["75.70.Kw","78.20.Ls","07.05.Mh","75.60.Ch"],"model":"grok-4.5","headline":"A single magneto-optical image of magnetic maze domains can reconstruct field, temperature, and magnetic history at once.","keywords":["magnetic textures","magnetic domains","magneto-optics","machine learning","deep learning","perpendicular magnetic anisotropy","multiparametric sensing","Bi:YIG"],"falsifier":"Train and test the same CNN and feature-vector models on an independent set of quantitative domain images taken from a different PMA film (different composition, thickness, or defect density) under the same field and temperature ranges; if residual field and temperature errors rise by more than a factor of a few and history classification fails, the claimed generality of the fingerprint collapses.","tokens_in":17674,"feed_emoji":"🧲","tokens_out":992,"duration_ms":10875,"temperature":0.7,"pith_summary":"Magnetic domain patterns are not just pretty pictures of magnetization. They are fingerprints of the energy landscape that formed them: applied field, temperature, and the path the material took through its hysteresis loop all leave distinct morphological marks. The authors show that a single quantitative polarization map of maze domains in a high-contrast Bi:YIG film contains enough information to recover all three parameters simultaneously. They do this both with a convolutional network that reads the full image and with a simpler network fed only twenty hand-crafted, physically interpretable features such as domain period, endpoint counts, and magneto-optical contrast. On held-out data the convolutional model reaches roughly 3.8 µT and 0.12 K residual accuracy while perfectly classifying the history branch. The feature-based model is only modestly worse, proving that most of the usable signal is already carried by a few measurable morphological quantities. The result turns magnetic texture into a multiparametric sensor and supplies a concrete list of which domain properties encode which external conditions.","feed_headline":"One domain image recovers field, temperature and magnetic history","feed_subtitle":"Maze patterns in Bi:YIG encode external conditions at microtesla and 0.1 K precision from a single map","key_machinery":"Dual inference pipeline: a seven-layer CNN operating on tiled, augmented polarization maps, paired with a multilayer perceptron that receives a 20-dimensional hand-crafted feature vector (mean and variance of magneto-optical rotation, Fourier-derived domain period, endpoint and branch counts of positive and negative domains, fractal dimension, Sobel gradients, etc.). The feature vector both enables accurate regression and identifies which morphological descriptors carry field, temperature, and history information.","core_discovery":"A single fine-scale, high-contrast magneto-optical polarization map of maze domains in epitaxial Bi:YIG encodes magnetic field, temperature, and hysteresis-branch history so completely that both a convolutional neural network and a 20-feature multilayer perceptron can reconstruct all three parameters from that one image, with the CNN reaching µ₀ΔH ≈ 3.8 µT and ΔT ≈ 0.12 K residual accuracy and perfect history classification.","pith_inferences":["Because the CNN residual is only modestly better than the feature vector, further gains may come more from richer physics-informed features than from deeper networks.","The observed decrease of normalized endpoint density with temperature suggests domain patterns become more crystalline at higher T; that entropy measure itself could become a diagnostic for anisotropy temperature dependence.","If history can be read so cleanly, multi-step field protocols should allow reconstruction of a coarser magnetic timeline rather than only the last branch.","The method supplies a practical benchmark dataset for testing whether inverse-inference networks trained on simulated Turing-type patterns transfer to real magnetic images."],"forward_implications":["Magnetic domain images become single-shot multiparametric sensors for simultaneous field, temperature, and magnetic history.","Hand-crafted descriptors already capture most of the usable signal, so lighter, interpretable models can replace black-box networks for many sensing tasks.","Endpoint density of one polarity relative to domain period encodes hysteresis branch even when average magnetization is identical.","Normalized residual sensitivities reach ~1.9 µT Hz^{-1/2} and ~0.06 K Hz^{-1/2}, competitive with dedicated multiparametric magneto-optical methods while adding history readout.","The same framework can be expanded to recover additional material parameters once they are varied systematically in the training data."],"fun_headline_variants":["Single domain map recovers field temperature and history","Maze patterns encode field temp and magnetic history","One polarization image decodes three magnetic parameters","Domain texture alone yields field temp and hysteresis branch","CNN and features reconstruct conditions from one maze map"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The high information density and the learned mappings demonstrated on this single low-defect commercial Bi:YIG film under quasi-static laboratory conditions will transfer to other materials, defect densities, or dynamic states without large loss of accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Single domain map recovers field temperature and history","Maze patterns encode field temp and magnetic history","One polarization image decodes three magnetic parameters","Domain texture alone yields field temp and hysteresis branch","CNN and features reconstruct conditions from one maze map"]},"model":"grok-4.5","effort":"low","cost_usd":0.003532,"raw_usage":{"total_tokens":1134,"prompt_tokens":770,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":35320000,"prompt_tokens_details":{"text_tokens":770,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":293,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":770,"tokens_out":71,"duration_ms":6486,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T18:16:12.282949+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and test the same CNN and feature-vector models on an independent set of quantitative domain images taken from a different PMA film (different composition, thickness, or defect density) under the same field and temperature ranges; if residual field and temperature errors rise by more than a factor of a few and history classification fails, the claimed generality of the fingerprint collapses.","supporting_citations":[],"review_version":2}