{"id":"b94af88f-4f84-433d-a39b-d3144b47ee9d","arxiv_id":"2607.29249","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Machine-learning surrogates and two-constant symbolic laws predict coercive field, remanence, and energy product of an ideal hard-magnet grain with far lower held-out error than Stoner–Wohlfarth and Kronmüller models.","lead":"Using 12,012 micromagnetic simulations of an ideal 50-nm grain, the authors train machine-learning surrogates and two-constant symbolic laws that map material constants to coercive field, remanence, and energy product, and show the exchange constant cannot be recovered back from those outputs. A generalist should read this as a well-benchmarked, honestly-scoped template for replacing expensive physics simulation with compact fitted laws — and for seeing where the replacement","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Target-derived filtering likely biases the symbolic-regression fits; testing the 1,624 excluded non-reversing samples against Eq. (9) would settle whether the laws generalize.","rationale":"The reader's weakest assumption identifies the target-derived population as the core limitation. I agree and sharpen it: the exclusion of non-reversing samples truncates Hc, and the hard/soft k-means uses Mr/Ms, so the symbolic-regression fits and reported RMSEs are conditional on the target. Since the data availability statement promises the simulation dataset, the invalid samples' (Ms, A, K) should be available, allowing a direct consistency check of Eq. (9). If the law predicts Hc above the reversal threshold for most invalid samples, the filtering is benign for that law; if not, the law and error numbers are artifacts of the filter. This is the single most load-bearing concern because it affects the core claim that the symbolic laws are accurate surrogates for the studied physics. The size ansatz is secondary and explicitly acknowledged. The paper is honest and reproducible, so no verdict change beyond the reader's CONDITIONAL is warranted; the proposed test could strengthen or weaken confidence.","tokens_in":15047,"tokens_out":10071,"duration_ms":96533,"concrete_test":"For each of the 1,624 invalid (non-reversing) samples, compute the Eq. (9) prediction Hc_pred using the released (Ms, A, K) and the fitted α = 0.942, n = 0.0921. If the law is unbiased, the fraction of invalid samples with Hc_pred < 7.96 MA/m (the field at which reversal should have been observed) should be small, because these samples did not reverse by −10 T. If a large fraction (say >20%) have Hc_pred < 7.96 MA/m, the law is inconsistent with the exclusion rule and the filtering has biased the fit. This test requires only the already-released invalid parameter sets, no new simulations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Eqs. (9)–(11) are closed-form laws approaching RF accuracy—is conditioned on a population selected on the target variables. Samples that did not reverse by μ0H = −10 T (1,624) are discarded; samples with Mr or Hc < 10^4 A/m are removed; and the 'hard-magnet subset' (8,831) is defined by k-means on (Ms, Mr/Ms), i.e., on a target. This selection couples the inputs (Sec. IV A) and truncates the Hc distribution at the applied-field maximum. The symbolic-regression search (Sec. IV E) therefore fits on a target-filtered set; the recovered form Hc = [α − n ln L̃/κ]HA may absorb selection bias rather than physics. The paper acknowledges the filtering but still presents the laws as 'new closed-form expressions.' The size-dependence is an ansatz, but the selection bias is more immediate: it affects the error numbers and the fitted constants even at the fixed 50-nm geometry. If the invalid samples are exactly those for which the law predicts Hc below the reversal threshold, the law is inconsistent with the data-generation mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains random-forest, Gaussian-process, and other ML surrogates on ~12,000 micromagnetic simulations of an isolated 50-nm cubic grain with uniaxial anisotropy, sampling Ms, A, and K over broad ranges. On a held-out test set restricted to a k-means-defined 'hard-magnet subset' (8,831 samples), the random forest and Gaussian process predict Hc, Mr, and BHmax with substantially lower RMSE than fitted analytical benchmarks (Kronmüller, Mr ≈ Ms, and μ0Ms²/4). Symbolic regression yields three compact closed-form laws, Eqs. (9)–(11), which the authors interpret as recovering the Kronmüller form for Hc with a material-dependent effective demagnetizing factor, and as new corrections for Mr and BHmax. The inverse problem is also studied: Ms and K are recovered well, while A is not. A released package, mammos-ai, implements the screening pipeline.","tokens_in":15193,"tokens_out":6679,"duration_ms":67257,"significance":"If the claims hold, this is a useful contribution to computational magnetism: it provides a reproducible, openly released ML surrogate that runs orders of magnitude faster than micromagnetics, and it proposes compact closed-form corrections that approach the accuracy of a random forest. The paper is also transparent in releasing data, code, and fixed-seed figure scripts, and it fits the Kronmüller α on training data only. The main risk, which the authors partly acknowledge, is that all headline results are conditioned on a target-derived population; this must be resolved before the symbolic laws can be regarded as general closed-form expressions.","major_comments":[{"comment":"The population on which the models are trained and evaluated is conditioned on the target variables: 1,624 non-reversing simulations (no reversal by μ0H = −10 T) are discarded, 431 samples below the Hc/Mr floors are removed, and the hard-magnet subset is defined by k-means on (Ms, Mr/Ms). Section IV A acknowledges that this filtering couples inputs to targets, but the symbolic-regression laws are nevertheless presented as 'new closed-form expressions' without testing on the excluded samples. A concrete and decisive test is to evaluate Eq. (9) on the 1,624 non-reversing samples: for these, the simulated Hc is beyond the applied-field range, so the law should predict Hc ≥ 10 T. If it does not, the recovered functional form and the reported RMSEs are artifacts of the selection. At minimum, the claims must be explicitly restricted to the target-conditioned hard-magnet population.","section":"§IV A, §IV E (Eqs. 9–11)"},{"comment":"The claimed recovery of the Kronmüller form is partly shaped by the dimensionless normalization Hc/HA, which is exactly the reduced variable used in the Kronmüller model. The expression Neff = 2n κ ln L̃ is a rewriting of the fitted expression rather than an independent measure, and the L-dependence is explicitly an ansatz (stated in §IV E). The paper should make this caveat more prominent and avoid implying that a validated size dependence has been discovered. In particular, Eqs. (9)–(11) apply only to the 50-nm cube geometry and to the conditioned population; their presentation as general closed-form laws is premature without additional geometry or excluded-sample checks.","section":"§IV E, Eq. (12)"},{"comment":"The hard/soft classification that defines the hard-magnet subset is based on k-means on (Ms, Mr/Ms)—i.e., on a target variable. While the comparison to the κ criterion in Fig. 2 is a legitimate classification test, the subsequent analysis inherits this target-derived definition. The hard-magnet subset is not the same as the set of materials satisfying the physical criterion Eq. (2): 385 hard magnets lie below the κ = 1/√6 line. The paper should clarify that the reported laws and error metrics apply to this cluster-defined subset, not to all materials that pass the analytical hardness criterion.","section":"§IV B"}],"minor_comments":[{"comment":"The table header lists 'Linear Regression (LP)' but the text and the table row use 'LR'. This should be made consistent.","section":"Table I"},{"comment":"The FCNN results for A are numerically absurd (MAPE 3.5 × 10^11 %, RMSE 7.8 × 10^9 pJ/m), which suggests a convergence or scaling problem rather than a meaningful model failure. This should be remarked on or the model excluded from the comparison.","section":"Table II, FCNN row"},{"comment":"The caption contains the fragment 'An invalid sample in Sec. III', which is incomplete and does not explain the figure. Please rewrite the caption to describe the pipeline flow.","section":"Fig. 9 caption"},{"comment":"The rounding/simplification step in the symbolic-regression procedure is described only verbally. For reproducibility, the original Pareto-front expression before exponent rounding should be reported, at least in the Supplemental Material.","section":"§IV E"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a condensed-matter/magnetism journal and the released pipeline is a genuine asset. The load-bearing issue is the target-conditioned population; I recommend asking the authors to perform the excluded-sample test (non-reversing samples against Eq. (9)) and to reframe the symbolic laws as conditional on the 50-nm cube and the filtered hard-magnet subset. If the excluded-sample test fails, the central claim of closed-form general laws would need substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent and unusually honest ML-for-micromagnetics paper. It builds an open dataset of ~12k simulations of a 50-nm cube, benchmarks five ML models against the standard analytical estimates on identical held-out data, and shows RF/GP beat Kronmüller by a factor of 2–10 in RMSE. The trained models, ONNX weights, and per-figure reproduction scripts are released. That alone earns a serious referee.\n\nWhat's new: the dataset and the mammos-ai pipeline; the explicit closed-form corrections, Eqs. (9)–(11), with at most two fitted constants, approaching RF accuracy; and the negative inverse-problem result for A, which is credible. The symbolic regression recovering the Kronmüller form, with Neff = 2nκ ln L̃, is a nice reframing but it is a rediscovery of a known form—the authors themselves cite ref. [31] for the logarithmic dependence. Eq. (11) is the square of Eq. (10) by their own statement. So the \"new laws\" claim is narrower than the abstract suggests, but not empty.\n\nSoft spots in proportion. First, all headline numbers are conditioned on a target-derived population: non-reversing simulations are discarded, low-Mr / low-Hc samples are removed, and the \"hard-magnet subset\" is k-means clustered on Mr/Ms. The paper acknowledges this in Sec. IV A, and for a hard-magnet screening tool the conditioning is defensible, but it means the fitted constants and error bars are not representative of the full stated input ranges. The stress-test concern about the symbolic laws absorbing selection bias is worth taking seriously. The obvious check—testing Eq. (9) on the 1,624 non-reversing samples to see whether the law extrapolates or breaks—is not done, and the authors instead present the laws as generally applicable. Minor but real: Table I shows the GP is about 5× better than the RF on Hc (18 vs 92 kA/m RMSE), yet the prose calls the accuracy gap \"negligible.\" That is an overstatement. The size-dependence is explicitly an ansatz, so the \"laws\" are fixed-geometry fits, not validated scaling relations.\n\nBottom line: this is a well-scoped, transparent study. The central claim—that ML surrogates beat analytical benchmarks on this dataset—holds. The symbolic laws are less novel than advertised, and the selection-bias caveat caps their generality, but the paper discloses most of this itself. It is for anyone doing micromagnetics screening or building surrogate models in the MaMMoS ecosystem. I would send it to review, with a request to correct the \"negligible\" language and ideally add the non-reversing-sample test if the data allow.","headline":"A useful, honest surrogate-modeling paper for ideal hard-magnet grains; the ML gains are real, the symbolic laws are partly rediscoveries with an unvalidated size ansatz, and the headline errors apply only to a target-conditioned subset.","tokens_in":15914,"tokens_out":2057,"would_cite":true,"duration_ms":20606,"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":"Machine-learning surrogates and symbolic regression can replace expensive micromagnetic simulations for predicting hard-magnet hysteresis properties, cutting held-out coercive-field error from 195 to 18 kA/m while recovering the Kronmüller","keywords":["micromagnetic simulation","hysteresis modelling","hard magnets","coercive field","symbolic regression","machine-learning surrogate","inverse design","permanent magnets"],"falsifier":"Run the same symbolic-regression search on the full unfiltered population, or on cubes of side 100 nm and 200 nm with identical sampling; if the fitted exponents of L̃ in Eqs. (10)–(11) shift beyond the reported standard errors, the recovered laws are artefacts of the 50-nm conditioning rather than general relations.","tokens_in":14742,"feed_emoji":"🧲","tokens_out":8589,"duration_ms":71661,"temperature":0.7,"pith_summary":"The paper tries to establish that the expensive step in permanent-magnet modelling—computing hysteresis properties from micromagnetic parameters—can be replaced by fast surrogate models and, more strongly, by compact closed-form laws. Using roughly twelve thousand micromagnetic simulations of an idealized 50-nm cube as ground truth, it shows that random-forest and Gaussian-process surrogates predict coercive field, remanence, and maximum energy product with substantially lower held-out error than the standard analytical models. Symbolic regression then recovers the Kronmüller form of coercivity, with an effective demagnetising factor that depends on the material, and produces new closed-form expressions for remanence and energy product with at most two fitted constants each. The practical payoff is that thousands of candidate parameter sets can be screened in seconds rather than hours. The paper is careful to state that these results are conditional on the hard-magnet subset and the 50-nm geometry.","feed_headline":"Hard-magnet error drops from 195 to 18 kA/m","feed_subtitle":"Learned surrogates and two-constant formulas turn hours of micromagnetic simulation into second-scale screening.","key_machinery":"The analysis is built on two dimensionless combinations of the intrinsic parameters: the hardness parameter κ = sqrt(K / (µ0 Ms²)) and the reduced grain size L̃ = L / ℓex, with ℓex the exchange length. Scaling each extrinsic target by its natural analytical unit (the anisotropy field HA for Hc, Ms for Mr, µ0Ms²/4 for BHmax) lets symbolic regression search a compact dimensionless space; the key output is the effective demagnetising factor Neff = 2n κ ln L̃, which converts the fitted constant in the Kronmüller coercivity formula into a material-dependent quantity. The same construction yields correction terms in the Mr and BHmax laws that grow when the grain spans many exchange lengths.","core_discovery":"The paper establishes that, for an idealised 50-nm cubic hard-magnet grain, the mapping from the three intrinsic constants (saturation magnetisation Ms, exchange stiffness A, uniaxial anisotropy K) to the three extrinsic hysteresis properties (coercive field Hc, remanence Mr, maximum energy product BHmax) is learnable and compressible. Trained on 12,012 micromagnetic simulations, random-forest and Gaussian-process surrogates predict all three extrinsic quantities on held-out data with substantially lower error than the best analytical benchmarks: Hc RMSE of 92 and 18 kA/m versus 195 kA/m for the fitted Kronmüller form, Mr RMSE of 11 and 4 versus 53 kA/m, and BHmax RMSE of 17 and 2 versus 148","pith_inferences":["If the same dimensionless construction is applied to other grain shapes, the correction structure may carry over, giving a template for shape-agnostic hard-magnet surrogates.","The near-zero influence of A on the extrinsic properties implies that inverse design from hysteresis alone is fundamentally underdetermined; a complementary measurement that couples to exchange, such as ferromagnetic-resonance frequency or nucleation-field angular dependence, would be needed to pin A.","The paper's own caveat that the size dependence is an ansatz suggests a direct multi-size test: if the fitted exponents of L̃ change across cube sizes, the formulas are geometry-conditioned fits rather than general laws."],"forward_implications":["A material scientist can evaluate a candidate (Ms, A, K) triple in about a second per material and batch-screen 10,000 candidates in under three seconds, compared with a median 147 minutes per micromagnetic simulation.","For the four published hard-magnet parameter sets tested, both the released pipeline and the symbolic laws reproduce dedicated simulations to within 12% for coercivity, 5% for energy product, and 3% for remanence.","The recovered coercivity law is the Kronmüller form with Neff = 2nκ ln L̃, so the old fitted constant becomes a material-dependent quantity.","The inverse models recover Ms and K from Hc, Mr, and BHmax accurately, but A cannot be recovered from these three outputs, making the inverse problem underdetermined for exchange stiffness.","The analytical hardness criterion κ > 1/√6 classifies 96.3% of simulated grains as hard or soft, with errors only in the conservative direction."],"fun_headline_variants":["ML and two-constant laws outpredict magnet models","Hard-magnet error cut 10x by learned surrogates","Symbolic regression finds simple laws for magnet properties","Seconds-scale magnet screening via ML surrogates","Predicting hard-magnet properties without micromagnetics"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the reported accuracies and formulas describe hard-magnet grains in general; in fact they are conditional on the paper's own filtering (reversal within −10 T, Hc and Mr above 10⁴ A/m, length scales above the 1-nm mesh, and a k-means hard/soft split) and on a single 50-nm cube with easy-axis field at zero temperature, a point the paper itself flags in Sections IV A and IV E.","fun_headline_variants_meta":{"raw":{"variants":["ML and two-constant laws outpredict magnet models","Hard-magnet error cut 10x by learned surrogates","Symbolic regression finds simple laws for magnet properties","Seconds-scale magnet screening via ML surrogates","Predicting hard-magnet properties without micromagnetics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1218,"prompt_tokens":820,"completion_tokens":398,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":321}},"tokens_in":564,"tokens_out":398,"duration_ms":4432,"temperature":1.0,"reasoning_tokens":321,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:48:41.673874+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same symbolic-regression search on the full unfiltered population, or on cubes of side 100 nm and 200 nm with identical sampling; if the fitted exponents of L̃ in Eqs. (10)–(11) shift beyond the reported standard errors, the recovered laws are artefacts of the 50-nm conditioning rather than general relations.","supporting_citations":[],"review_version":1}