{"id":"438a1730-545f-4094-9936-ad7802695811","arxiv_id":"2501.06171","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A machine-learning force field with symmetry-preserving descriptors reproduces non-collinear spin dynamics and reveals arrested skyrmion ordering in triangular-lattice s-d models.","lead":"This paper trains neural networks to compute the magnetic forces that drive spin dynamics in metals where electrons and local magnetic moments interact. The method reproduces known magnetic orders and suggests that skyrmion crystals freeze into glassy stripe patterns instead of ordering when quenched.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The rc=6a locality cutoff equals the skyrmion ordering period, and since single-/double-/triple-Q states are nearly degenerate, omitted RKKY tails could bias the ML landscape and manufacture the arrested glassy state.","rationale":"The reader's weakest_assumption identifies exactly the locality cutoff as the load-bearing risk. My reading of the manuscript confirms that the central physical claim is the arrested glassy skyrmion state, which is obtained only from ML-LLG at 150x150 and not checked against exact KPM-LLG at comparable size. The paper's own statement that the competing Q states are nearly degenerate makes the truncation at one ordering period especially dangerous: even small missing long-range contributions can reorder the energy landscape. The proposed energy comparison directly tests whether the ML landscape has the correct minima; if it does not, the dynamics are unreliable. This does not change the reader's conditional verdict, since the paper should address exactly this concern, for example by convergence in rc or exact validation on the largest feasible system.","tokens_in":97,"tokens_out":7912,"duration_ms":146276,"concrete_test":"Compute the exact KPM energy per site of the single-Q helix, double-Q, and triple-Q skyrmion-crystal states on a commensurate large lattice (e.g., 72x72 or 150x150) and compare with the energies predicted by the ML model. If the ML model does not reproduce the exact ordering (triple-Q lowest) to within the training error, or if the energy differences are smaller than the ML energy error, the rc=6a truncation biases the landscape and the Sec. V arrested state is suspect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central new physical result is the arrested phase ordering and glassy skyrmion state of Section V. The ML model assumes a local energy depending only on spins within rc=6a (Eq. (7); Section IV). For the skyrmion s-d model, the ordering wavevector is Q=(pi/3a,0) (and rotations), so the magnetic period is exactly 6a, and electron-mediated interactions are long-ranged RKKY-like (Eq. (6) and surrounding text). The paper itself states that the single-, double-, and triple-Q states are 'nearly degenerate in energy' (Section V). If interactions beyond 6a contribute at the level of this near-degeneracy, the ML landscape can favor single-Q stripes/bimerons over the triple-Q skyrmion crystal, making the arrest an artifact. The small training MSE only shows the local model fits sampled 48x48 configurations; it does not establish that truncation is harmless for large-scale coherent textures or for late-time stripe states, and no exact large-scale validation is reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a Behler-Parrinello-type machine-learned force field for adiabatic Landau-Lifshitz-Gilbert (LLG) dynamics of s-d itinerant electron magnets. The proposed magnetic descriptors decompose bond and chirality variables into irreducible representations (IRs) of the lattice point group and use a reference-IR construction to recover phase information lost in a power-spectrum representation. The authors train feedforward neural networks on KPM-computed local fields for three triangular-lattice s-d models (120-degree order, tetrahedral order, and a triple-Q skyrmion crystal), report torque MSEs on the order of 1e-7, and use the ML surrogate to perform 150x150 thermal quenches. The central new physics claim is that the skyrmion crystal exhibits arrested phase ordering: the system freezes into stripe and bimeron textures rather than forming a large coherent triple-Q skyrmion crystal.","tokens_in":52596,"tokens_out":4977,"duration_ms":53834,"significance":"If the central result holds, the paper is a useful demonstration that ML force fields can extend spin-dynamics simulations of itinerant magnets to system sizes inaccessible to direct KPM or ED, and the reference-IR descriptor construction is a meaningful methodological contribution. The paper ships concrete benchmarks: held-out torque MSEs, a KPM-vs-ML comparison of time-dependent structure factors on 48x48 lattices (Fig. 5), and a stability check of the ML skyrmion crystal under perturbation (Fig. 6). These support accuracy for in-distribution 48x48 dynamics. However, the headline physical result depends on the ML energy landscape being faithful for late-time 150x150 textures whose ordering period equals the locality cutoff, and on the model resolving the near-degeneracy among single-, double-, and triple-Q states. That load-bearing point is not yet benchmarked, so the significance of the central claim is currently conditional.","major_comments":[{"comment":"The locality cutoff rc=6a is commensurate with the magnetic period of the skyrmion ordering: the ordering wave vectors are Q1=(pi/3a,0) and rotations, so the period along the ordering direction is exactly 6a. The effective spin interactions in s-d models are RKKY-like and long-ranged (Eq. (6) and surrounding text), and Section V states that the single-, double-, and triple-Q states are nearly degenerate in energy. Neither the torque MSE in Fig. 3 nor the 48x48 structure-factor comparison in Fig. 5 establishes that omitting spin correlations beyond rc=6a is harmless for the 150x150 late-time stripe and bimeron configurations. Because the ordering period equals the cutoff, the truncation could bias the relative stability of the competing Q states and manufacture the observed arrested glassy behavior. Please provide a quantitative test, such as a KPM-vs-ML comparison of M(t) and L(t) on the largest feasible lattice (96x96 or 120x120) at late times, and an estimate of the energy contribution from spins beyond rc for representative single-, double-, triple-Q, and late-time stripe states.","section":"Section IV.C / Section V"},{"comment":"The paper reports that 'using high precision KPM, we found that the single, double, and triple-Q states are nearly degenerate in energy, with a slightly lowered energy for the triple-Q skyrmion crystal,' but it gives no numerical values for these energy differences and no ML prediction of them. The training loss is dominated by torque MSEs of order 1e-7, but no energy error or energy ranking of the competing states is reported. If the energy differences controlling the Q-state competition are smaller than the ML energy error, the ML landscape can invert the ordering and the arrested phase ordering would be an artifact. Please report the KPM and ML energies for single-Q, double-Q, triple-Q, and the late-time stripe/bimeron configurations, and show that the ML model ranks them in the same order as the exact KPM energies.","section":"Section V"},{"comment":"The dataset provenance is stated inconsistently: the text says the 40,000 configurations were obtained 'from 40 independent ED-LLG simulations' and then says 'The training dataset was obtained from KPM-based LLG simulations on a relatively small 48x48 lattice.' These are different solvers. This matters for reproducibility and for assessing whether the ML model learns exact or approximate local fields. Please clarify which solver generated the training data and whether ED and KPM were cross-checked for the training configurations.","section":"Section IV.A"}],"minor_comments":[{"comment":"The heading 'T riple-Q Skyrmion lattice' and the phrase 'tripe-Q structure' in Fig. 3(i) contain typos and should be corrected.","section":"Section IV.C / Fig. 3 caption"},{"comment":"In Eq. (34), the first component reads 'cos Q1i - 1/2 cos Q2i - 1/2 Q3i'; the last term is missing the cosine and should presumably be '-1/2 cos Q3i'.","section":"Eq. (34)"},{"comment":"Eq. (7) defines the neighborhood Ci by a hard cutoff rc, while Eq. (11) introduces a soft cutoff function fc(r) with a smoothing width. Please clarify whether the hard and soft cutoffs are the same radius and whether the soft cutoff is used in the final feature construction.","section":"Section II.B"},{"comment":"The definition of the six symmetry-related blocks used for the reference IR is not precise: the caption says the blocks 'can partially overlap with each other,' but the assignment rule and the averaging procedure should be specified in the text, because the reference basis depends on this choice.","section":"Section III.D / Fig. 2(d)"},{"comment":"The paper would benefit from a data/code availability statement. Given the methodological nature of the work, making the descriptor and training code available would substantially improve reproducibility.","section":"General"},{"comment":"The relation to prior work by the same group (Refs. [41], [43], [47]) is described only by citations; a short explicit sentence distinguishing the new contribution of this paper (the reference-IR descriptor demonstration with the arrested-ordering result) would help the reader understand the incremental advance.","section":"Introduction / Section VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is a hybrid review/methods article. The reference-IR descriptor concept appears in earlier work by the same group (Ref. [41]), and the tetrahedral chiral-domain dynamics appears in Ref. [47]; the editors may wish to ask the authors to sharpen the statement of novelty relative to those works. There is no code/data availability statement, which is a growing expectation for ML-methods papers. The major technical risk is the rc=6a locality cutoff for a system whose ordering period is also 6a; the authors should be encouraged to provide the largest-scale exact benchmark they can afford, even if it is not the full 150x150 system."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read the paper with the stress-test concern in mind. My take: the reader's conditional verdict is right, and the locality concern is real, not manufactured.\n\nWhat is genuinely new is the application, not the machinery. The reference-IR descriptor is explicitly from earlier work, and the s-d benchmarks extend prior same-group papers. The new content is the systematic benchmark across three phases and the 150x150 quench result: instead of a coherent skyrmion crystal, the system freezes in stripe/bimeron states. That is a concrete, testable scenario and a legitimate reason to read the paper.\n\nWhat it does well: the benchmarks are the right kind. Torque MSEs around 1e-7 to 1e-6 on held-out configurations, and the KPM-vs-ML structure factor comparison in Fig. 5 is a real dynamical check, not just a fit to energy. The symmetry discussion is careful, and the paper is honest that the single/double/triple-Q states are nearly degenerate.\n\nThe soft spot is the locality assumption. With rc=6a and Q=(pi/3a,0), the cutoff equals one magnetic period. The effective exchange in a metal is RKKY-like with long oscillatory tails, and the energy differences among competing Q states are small by the paper's own account. A model that cannot see interactions beyond one period can plausibly bias the landscape toward single-Q stripes and manufacture the arrested glassy state. The small training MSE and the 48x48 structure factor do not address this: the late-time stripe states are not in the training distribution, and no exact large-scale check is reported. I would want a direct test: retrain with rc=8a or 10a, and compare ML versus KPM energy differences for single-, double-, and triple-Q configurations. Providing code and data would also help; their absence is a genuine weakness for a ML paper.\n\nI don't share the circularity objection: training on exact data from the same model is normal, and the test is dynamical reproduction.\n\nBottom line: the method is credible and the paper is worth a serious referee. I would send it to peer review but insist on cutoff-sensitivity and exact-energy checks before the arrested-ordering claim is accepted. I'd cite the method for its benchmarks.","headline":"Careful ML-LLG benchmarks on top of an established descriptor, but the headline skyrmion-freezing claim rests on a locality cutoff that needs a direct test before I'd trust it.","tokens_in":53083,"tokens_out":3687,"would_cite":true,"duration_ms":40312,"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 machine-learned force field for spin dynamics predicts local fields in itinerant magnets, reproduces three non-collinear orders, and shows skyrmion crystals freezing into a glassy stripe state.","keywords":["machine learning force fields","Landau-Lifshitz-Gilbert dynamics","skyrmion crystal","itinerant electron magnets","s-d model","symmetry-invariant descriptors","reference irreducible representations","glassy skyrmion state"],"falsifier":"Retrain the model with a larger cutoff radius, for example $r_c = 10a$, and repeat the 150x150 thermal quench; if the frozen stripe state and plateaued triple-Q order disappear, the glassy state is a truncation artifact rather than a physical property.","tokens_in":52157,"feed_emoji":"🧲","tokens_out":5105,"duration_ms":49320,"temperature":0.7,"pith_summary":"The paper claims that a machine-learned force field can replace the expensive quantum-mechanical calculation of local exchange fields in itinerant electron magnets, making large-scale spin-dynamics simulations practical. The approach assumes the effective energy of a spin depends only on its immediate neighborhood, and encodes that neighborhood with descriptors invariant under both global spin rotations and the lattice point group. Trained on exact data from the triangular-lattice s-d model, the method reproduces 120-degree, tetrahedral, and skyrmion-crystal orders. In 150x150 thermal quenches, the skyrmion crystal fails to order globally and freezes into a glassy stripe state of skyrmions and bimerons.","feed_headline":"Machine-learned spin forces reveal a glassy skyrmion state","feed_subtitle":"Symmetry-invariant descriptors let spin dynamics reach 150x150 lattices and expose arrested ordering.","key_machinery":"The load-bearing object is the magnetic descriptor: a mapping from the local spin configuration $C_i$ to invariant feature variables. Starting from bond variables $b_{jk} = \\mathbf{S}_j \\cdot \\mathbf{S}_k$ and scalar chiralities $\\chi_{jmn} = \\mathbf{S}_j \\cdot \\mathbf{S}_m \\times \\mathbf{S}_n$, the paper decomposes them into irreducible representations of the $D_6$ site-symmetry group and uses reference irreducible representations to supply the phase information missing from a power-spectrum descriptor. The resulting features are invariant under global spin rotations and lattice point-group operations, differentiable with respect to spin rotations, and are fed into a feedforward neural network that outputs the local energy $\\epsilon_i$; automatic differentiation then yields the local field $\\mathbf{H}_i = -\\partial E/\\partial \\mathbf{S}_i$.","core_discovery":"The central claim is that a machine-learning force field of the local-energy type, built from reference irreducible representations of the site-symmetry group, gives local effective fields accurate enough to drive faithful Landau-Lifshitz-Gilbert dynamics for the s-d model. Because the local energy is a symmetry-invariant function of bond and scalar-chirality variables in a cutoff neighborhood, the predicted torques respect the model's global spin-rotation and lattice point-group symmetries by construction. The trained models reproduce the 120-degree, tetrahedral, and triple-Q skyrmion-crystal phases of the triangular-lattice s-d model, with torque errors around $10^{-7}$. Large-scale quenches then reveal that skyrmion crystallization is arrested: the system freezes into disordered stripe structures containing skyrmions and bimerons, with a ring-like structure factor and a plateaued triple-Q order parameter, which the paper interprets as a glassy skyrmion state.","pith_inferences":["The frozen stripe state may be sensitive to the locality cutoff: skyrmion order in s-d models is mediated by long-range oscillatory electron-mediated interactions, so correlations beyond the $6a$ cutoff could in principle change the balance between single-Q and triple-Q states.","A direct test is to retrain with increasing cutoff and check whether the arrested phase and glassy state persist; if they vanish, the glass is a truncation artifact.","The descriptor construction from bond and chirality variables should transfer to other two-dimensional Bravais lattices and to three-dimensional magnets, where the relevant point groups have larger irreducible representations.","The method's ability to reach large systems cheaply makes it a candidate for scanning parameter space of frustrated itinerant models for metastable topological textures."],"forward_implications":["LLG simulations using only ML-predicted local fields reproduce the 120-degree, tetrahedral, and skyrmion-crystal orders, so the energy landscape learned from local neighborhoods is sufficient for these ordered phases.","The quench dynamics of the skyrmion crystal show arrested growth of the triple-Q order and a ring-like structure factor, implying the system gets stuck in metastable stripe states rather than crystallizing.","The same framework extends to systems with spin-orbit coupling by replacing the separate spin-rotation and lattice symmetries with a combined spin-lattice symmetry group.","Because the local field is computed from a fixed-size neural network, simulation cost scales linearly with system size, enabling 150x150 lattices that would be prohibitive with repeated exact diagonalization."],"supporting_citations":[{"why":"Introduces the local-energy neural-network architecture that the paper adapts from molecular dynamics to spin systems.","marker":"[24]"},{"why":"Introduces the Gaussian approximation potential and bispectrum-style descriptor approach that the reference-IR method modifies.","marker":"[25]"},{"why":"Proposes the reference irreducible-representation construction that the paper uses to build symmetry-invariant magnetic descriptors.","marker":"[41]"},{"why":"Provides the kernel polynomial method used to generate exact training data and KPM-LLG benchmark trajectories.","marker":"[61]"},{"why":"Identifies the triple-Q and tetrahedral orders of the triangular-lattice s-d model that serve as the paper's benchmark phases.","marker":"[14]"},{"why":"Gives the specific tight-binding model with quasi-nesting that stabilizes the skyrmion crystal studied in Section IV.C.","marker":"[87]"},{"why":"Explains the multiple-spin interactions that stabilize the noncoplanar tetrahedral order used as a benchmark.","marker":"[60]"},{"why":"Previous machine-learning force-field study of the same models whose chiral-domain coarsening results are compared with.","marker":"[47]"}],"fun_headline_variants":["Machine-learned spin forces show glassy skyrmion freeze","AI force fields expose frozen skyrmion stripes","Symmetry-aware ML reproduces complex spin orders","ML spin dynamics reveal arrested skyrmion glass","Machine learning captures spin frustration and glass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The local energy of a spin is assumed to depend only on spins within roughly six lattice constants, while the electron-mediated interactions that stabilize skyrmion order are long-ranged and oscillatory.","fun_headline_variants_meta":{"raw":{"variants":["Machine-learned spin forces show glassy skyrmion freeze","AI force fields expose frozen skyrmion stripes","Symmetry-aware ML reproduces complex spin orders","ML spin dynamics reveal arrested skyrmion glass","Machine learning captures spin frustration and glass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000493,"raw_usage":{"total_tokens":2419,"prompt_tokens":943,"completion_tokens":1476,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":1418}},"tokens_in":559,"tokens_out":1476,"duration_ms":9726,"temperature":1.0,"reasoning_tokens":1418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:20.513068+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the model with a larger cutoff radius, for example $r_c = 10a$, and repeat the 150x150 thermal quench; if the frozen stripe state and plateaued triple-Q order disappear, the glassy state is a truncation artifact rather than a physical property.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Gaussian approximation potential and bispectrum-style descriptor approach that the reference-IR method modifies."},{"cited_title":"Ozawa, S","cited_arxiv_id":null,"evidence_quote":"Gives the specific tight-binding model with quasi-nesting that stabilizes the skyrmion crystal studied in Section IV.C."},{"cited_title":"Akagi, M","cited_arxiv_id":null,"evidence_quote":"Explains the multiple-spin interactions that stabilize the noncoplanar tetrahedral order used as a benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Previous machine-learning force-field study of the same models whose chiral-domain coarsening results are compared with."}],"review_version":1}