{"id":"f5cedd78-4560-4054-9c5c-6bacf0b6fde6","arxiv_id":"2411.19780","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN trained on exact diagonalization predicts local spin-flip energies and enables large-scale kinetic Monte Carlo simulations that reveal temperature-dependent domain growth in a double-exchange Ising magnet.","lead":"Researchers trained a convolutional neural network to predict the energy cost of flipping a single Ising spin in a double-exchange magnet, replacing expensive electron calculations with a fast local estimate. They used it to run large kinetic Monte Carlo simulations and report that magnetic domains coarsen unusually slowly at very low temperature, with growth exponent dropping from 1/2 to 1/4.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 21x21 locality cutoff is unvalidated in the large-domain low-temperature regime where the alpha=1/4 claim lives; the ED-kMC benchmark only covers early times on 30x30 lattices, so the surrogate may be driving the anomalous exponent.","rationale":"The reader identifies the locality cutoff and the out-of-distribution training manifold as the weakest assumption; I agree, and I have sharpened the concern by noting why the 21x21 cutoff is especially dangerous for the T=0.01 coarsening claim: interior spins in large ferromagnetic domains have a saturated local input, so the CNN cannot represent any dependence of the flip energy on the distant domain boundary. The 30x30 ED-kMC benchmark in Fig. 4 only reaches nstep=10000, corresponding to t ~ 11 in scaled time, whereas the alpha=1/4 exponent is extracted from long-time L(t) data on a 100x100 lattice (Fig. 6) and from late-time snapshots on 200x200 (Fig. 5). The surrogate error sigma=0.036 in Delta E is exponentially amplified at T=0.01, so even a small bias in interior-flip energies can change the acceptance probabilities and, consequently, the domain-growth exponent. The proposed test directly checks whether Delta E for an interior spin varies with domain size; if it does not, the locality assumption is adequate in the relevant regime and the concern is resolved. This does not change the reader's CONDITIONAL verdict, since the requested validation is still missing, but it does not strengthen the verdict beyond conditional because the methodology is plausible and the early-time ED-kMC agreement is evidence in favor of the surrogate.","tokens_in":52129,"tokens_out":4513,"duration_ms":44550,"concrete_test":"Use KPM (or ED for L<=60) to compute the exact Delta E for flipping a spin at the center of a perfect ferromagnetic square domain of linear size L=30, 60, 100, 200 on a lattice with periodic boundary conditions, while keeping the 21x21 input patch fixed. If Delta E(L) changes by more than about 0.05 t_nn across this range, the locality assumption Eq. (6) fails in the coarsening regime and the T=0.01 exponent from the CNN is not a reliable property of the model. As a complementary check, retrain the CNN with a 31x31 (and, if feasible with KPM-generated training data, 41x41) neighborhood, split the data by original snapshot before augmentation, and rerun the 100x100 T=0.01 quench; if the extracted alpha moves outside 0.25 +/- 0.03, the original claim is an artifact of the cutoff.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline physics, the alpha=1/4 coarsening at T=0.01 in Sec. IV, rests entirely on the CNN surrogate in a regime where the surrogate's validity is not established. Equation (6) assumes h_i depends only on a 21x21 neighborhood. In the large-JH limit of the Ising-DE model, electrons are confined to ferromagnetic clusters, and the kinetic-energy gain of a cluster is a nonlocal functional of cluster size and shape. An interior spin in a 100x100 domain has a 21x21 input that is essentially all +1; the CNN assigns it a fixed Delta E, while the exact Delta E can depend on domain size through the finite-size electron spectrum. This is precisely the straight-wall, island-dominated regime shown in Fig. 5 at T=0.01. The only dynamical validation, Fig. 4, compares ED and ML correlation functions on 30x30 lattices at nstep <= 10000 (t ~ 11); at these early times domains are smaller than the 21x21 cutoff, so the benchmark does not exercise the extrapolation to large domains. No convergence test with a larger neighborhood is shown, and the 100 training snapshots cannot be assumed to cover large ordered domains. Because acceptance probabilities enter as exp(-Delta E/T), a surrogate error of the size seen in Fig. 2 (sigma = 0.036 in Delta E) is amplified by roughly exp(Delta E/0.01) at T=0.01. A systematic bias in interior-flip energies as domains grow would directly change coarsening kinetics and could produce a spurious alpha=1/4.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a convolutional-neural-network surrogate for the local effective field h_i of an itinerant Ising double-exchange model, assuming that h_i depends only on spins in a fixed 21×21 neighborhood (Eq. (6)). The network is trained on exact-diagonalization (ED) energy differences computed on 30×30 lattices and then used inside kinetic Monte Carlo (kMC) simulations of much larger systems. The authors benchmark the surrogate against ED for single-flip energies, against equilibrium Monte Carlo for the ferromagnetic transition temperature, and against ED-kMC for the spin-spin correlation function after a quench to T=0.01. They then use the ML-kMC method on 100×100 and 200×200 lattices and report a temperature-dependent coarsening exponent: L(t)~t^{1/2} at T=0.1 but L(t)~t^{1/4} at T=0.01, with the exponent tending to zero as T→0. The central claims are that the CNN provides a transferable, linear-scaling energy model for kMC and that the Ising double-exchange model exhibits anomalous low-temperature phase-ordering kinetics.","tokens_in":52475,"tokens_out":5889,"duration_ms":61306,"significance":"If the locality assumption is valid and the surrogate errors are controlled, the method would be a useful discrete-variable analogue of ML force fields, enabling large-scale kMC simulations of itinerant systems whose energy updates are normally prohibitively expensive. The paper includes genuine external checks: the ED-kMC dynamical comparison on 30×30 lattices (Fig. 4) is a real test of the surrogate in a dynamical setting, the equilibrium Tc agrees well with the earlier result of Motome and Furukawa (ref. [48]), and the scaling collapse in Fig. 7 is a meaningful consistency test. The potential payoff is high because the reported anomalous coarsening at T=0.01, if correct, would be a new physical result for a disorder-free itinerant Ising model. However, the headline exponent currently rests on a regime in which the surrogate's validity is not established, and one of the reported accuracy metrics is compromised by the data-splitting procedure.","major_comments":[{"comment":"The test-set error is likely optimistic because the dataset is split after data augmentation at the level of individual augmented entries. The text states that each of the 100×900 original entries is expanded into 16 symmetry-related pairs and that the resulting 100×900×16 pairs are then split with a test fraction of about 15%. Since the 16 copies of one original entry share the same ΔE and are related by a symmetry transformation of the input, a random split at the augmented-entry level places equivalent copies in both training and test sets. The reported test MSE of 0.0014 and the width σ=0.036 in Fig. 2 therefore measure partly memorization rather than generalization. The split should be performed at the level of original (configuration, site) entries, or preferably at the level of whole snapshots, and the test metrics should be recomputed and reported.","section":"Sec. III, first paragraph and Fig. 2"},{"comment":"The locality assumption that h_i depends only on a 21×21 neighborhood is load-bearing and is not validated in the regime where the anomalous exponent is extracted. For the large-JH double-exchange model, the electron kinetic energy of a ferromagnetic cluster depends on the cluster as a whole through the finite-size electron spectrum; an interior spin in a large +1 domain has a 21×21 input that is almost entirely +1, so the CNN assigns that spin a fixed ΔE regardless of the actual domain size. The straight-wall and island morphologies shown in Fig. 5 at T=0.01 are exactly the configurations where this nonlocality should matter. The only dynamical benchmark, Fig. 4, is run on a 30×30 lattice for nstep≤10000, i.e. t≲11, where domains remain smaller than the cutoff; it does not exercise extrapolation to large ordered domains. Moreover, because acceptance probabilities enter as exp(−ΔE/T), an error of the size seen in Fig. 2 (σ=0.036) changes transition rates by factors of order exp(±3.6) at T=0.01, so a systematic bias in interior-flip energies could directly alter coarsening kinetics. The authors should provide a neighborhood-size convergence test (e.g., ℓ=15, 21, 31, 41) against ED on 30×30 or larger systems and a targeted test on large ordered/straight-wall configurations, or the α=1/4 result should be presented as tentative rather than established.","section":"Sec. II B, Eq. (6); Sec. IV and Fig. 5"},{"comment":"The representativeness of the training data is not documented. The text says only that '100 independent Ising configurations were collected' for the full dataset, with no information about the temperatures, ensembles, or equilibration protocols used to generate them, and no statement about whether the resulting configuration distribution includes the large ferromagnetic domains and straight interfaces that dominate the low-temperature kMC trajectories. Since the CNN is trained exclusively on these snapshots, the extrapolation to the late-time T=0.01 regime is an untested assumption. The authors should specify the snapshot-generation protocol and characterize the coverage of the training set relative to the configurations visited in the kMC simulations, for example by comparing distributions of domain sizes, energy barriers, and predicted ΔE values.","section":"Sec. III, first paragraph"},{"comment":"The evidence for the temperature-dependent exponent α(T) rests on two temperatures and two power-law fits, and the figure shows no error bars, no fitting windows, and no system-size comparison. The claim that α=1/4 is the correct asymptotic exponent for T=0.01 would be much stronger with a finite-size check (e.g., 100×100 vs 200×200 at overlapping times), a fit-window sensitivity analysis, confidence intervals on the fitted exponent, and at least one independent estimate of L(t) in the anomalous regime from a non-surrogate method (for example, KPM-based kMC on the largest feasible lattice, even if only for early times). The text also states that T→0 quenches freeze in disordered states 'based on both ED and ML', but no supporting data for this claim are shown; either the data should be presented or the claim should be softened.","section":"Sec. IV, Fig. 6 and Eq. (11)"}],"minor_comments":[{"comment":"The caption says the output node gives the local field h_i, while the main text in Sec. II B says the fully connected network 'outputs the ΔE_i energy difference'; Eq. (3) relates the two, but the text should state clearly which quantity is the direct network output.","section":"Fig. 1 caption and Sec. II B"},{"comment":"The sentence 'the run time of the ML model is independent of the system size' should be qualified as the inference time per spin-flip evaluation; the per-sweep cost is O(N), which is the actual linear-scaling statement.","section":"Sec. II B"},{"comment":"The manuscript does not report the CNN training details (optimizer, loss function, learning rate, number of epochs, batch size, number of trainable parameters, and the random seed used for the train/test split). These details are needed for reproducibility.","section":"Sec. III"},{"comment":"Fig. 6 would benefit from error bars on L(t) and an explicit statement of the time ranges used for the power-law fits; currently the reader cannot assess the statistical significance of the difference between α=1/2 and α=1/4.","section":"Sec. IV, Fig. 6"},{"comment":"The statement that quenches at T→0 become frozen 'based on both ED and ML' is not accompanied by any figure or quantitative description; please add the data or remove the unsupported claim.","section":"Sec. IV, end of section"},{"comment":"There are several minor grammatical and typographical issues, for example 'the ML force-field model offers' in Sec. II B should be 'The ML force-field model offers', and 'temperature is measured in unit of the nearest-neighbor hopping' in the Fig. 3 caption should read 'in units of t_nn'. A careful proofread would improve the presentation.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the methodological idea is timely. My main concern is not novelty but validation: the headline anomalous coarsening exponent at T=0.01 is produced entirely by a surrogate whose locality cutoff has not been converged and whose dynamical benchmark does not extend into the large-domain, straight-wall regime. The data-augmentation split issue is a separate but easily fixable problem that affects the reported test accuracy. I would be willing to support publication after the authors (i) fix the train/test split, (ii) add a neighborhood-size convergence study and a targeted test in the large-domain regime, and (iii) strengthen the exponent analysis with error bars and independent or finite-size checks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimately useful methods paper with a striking physics claim that outruns its validation. The CNN-kMC surrogate idea is clean, and the benchmarks are real: Tc matches ref [48], and the ED-kMC correlation functions on 30x30 agree. But the headline result — alpha=1/4 coarsening at T=0.01 — sits exactly in the regime the surrogate is least constrained, and the reported test MSE is likely optimistic because the train/test split appears to be done on the augmented data, so symmetry copies of the same local configuration may appear in both sets.\n\nThe new thing here is applying the Behler-Parrinello-style local-field idea to discrete kMC for itinerant magnets, with a CNN replacing the usual descriptor+NN. That is a natural extension of the group's earlier work, but a useful one. The paper is honest and well explained; the dynamical scaling collapse at both temperatures is a nice check.\n\nThe soft spots are three. First, the locality cutoff at 21x21 is asserted, not tested. In the large-JH limit electrons are confined to ferromagnetic clusters, so the energy of an interior spin flip depends on the cluster size and shape; a 21x21 window may be fine for small domains but is untested for the straight-wall and island textures that appear at T=0.01. The ED-kMC benchmark only reaches nstep~10000 on 30x30, when domains are still smaller than the cutoff, so it does not exercise the problematic regime. Second, the test MSE in Fig. 2 is compromised if the split is post-augmentation; the authors need to clarify and likely re-split by original configuration. Third, the anomalous exponent rests on two temperatures, with no error bars on the fit and no finite-size check; given that acceptance probabilities are exp(-DeltaE/T), a systematic bias in DeltaE at T=0.01 could produce exactly the sort of spurious slow growth reported.\n\nNone of this kills the paper. The method is plausible and the equilibrium benchmarks are genuine. But the physics claim should be treated as conditional until the cutoff is validated and the surrogate is shown stable against error injection. I would send it to a serious referee rather than desk reject — the referee can ask for the missing convergence test and a cleaner split — but I would not cite the coarsening result yet.\n\nWho for: people doing ML surrogates for discrete lattice dynamics, and anyone working on phase-ordering kinetics of double-exchange models. Worth a reading group if that is your crowd.","headline":"A promising CNN-kMC surrogate, but the headline low-T coarsening exponent rests on an unvalidated 21x21 locality cutoff and a possibly leaking train/test split.","tokens_in":53010,"tokens_out":4956,"would_cite":false,"duration_ms":45382,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.10.Ln","75.10.Hk"],"model":"deepseek-v4-flash","headline":"A CNN can learn the single-spin-flip energy of an itinerant Ising magnet, enabling linear-scaling kinetic Monte Carlo and revealing temperature-dependent coarsening that slows to $t^{1/4}$ at low temperature.","keywords":["machine learning force field","kinetic Monte Carlo","itinerant Ising model","double-exchange model","convolutional neural network","domain coarsening","Allen-Cahn law","locality principle"],"falsifier":"Compute the exact spin-flip energy by exact diagonalization for late-time $T=0.01$ snapshots of a $30\\times30$ system (straight walls and island domains) and compare with the CNN's prediction; if the typical error in $\\Delta E$ makes the acceptance probability $\\exp(-\\Delta E/T)$ deviate substantially from 1 at $T=0.01$, then the measured $\\alpha=1/4$ growth could be a surrogate artifact rather than a property of the model. A cleaner variant is to train a second network with a $31\\times31$ or $41\\times41$ neighborhood and check whether the fitted coarsening exponent changes.","tokens_in":51908,"feed_emoji":"🧲","tokens_out":6894,"duration_ms":56474,"temperature":0.7,"pith_summary":"The paper aims to make kinetic Monte Carlo simulations of itinerant Ising magnets feasible at large system sizes. In these models, the energy cost of flipping one spin is set by electrons that hop through the whole lattice, so every attempted flip normally requires an electronic-structure calculation. The paper argues that the local effective field at a site depends only on the spin configuration in a finite $21\\times21$ neighborhood, and that a convolutional neural network trained on exact-diagonalization data from a $30\\times30$ lattice can predict that field accurately. If this holds, the same trained network runs in constant time per spin flip, so a Monte Carlo sweep costs $O(N)$. Applying the network to quench dynamics of the Ising double-exchange model, the paper finds that ferromagnetic domains grow as $L(t)\\sim t^{1/2}$ at $T=0.1$ but only as $L(t)\\sim t^{1/4}$ at $T=0.01$, with a growth exponent that appears to vanish as $T\\to0$.","feed_headline":"Domain growth slows to t^{1/4} in an itinerant magnet","feed_subtitle":"A CNN energy surrogate makes large-scale Monte Carlo feasible and exposes temperature-dependent coarsening.","key_machinery":"The load-bearing object is the locality assumption expressed in Eq. (6): $h_i = F(\\{\\sigma_j : j \\in B_i\\})$, where $B_i$ is a $21\\times21$ square block centered at site $i$ and $F$ is a universal function learned by a CNN. The CNN (seven convolutional layers with a $7\\times7$ kernel followed by $3\\times3$ kernels, then a fully connected layer with 64 nodes and a single output node) predicts $\\Delta E_i$; because the kernel size is fixed, each evaluation is $O(1)$ and a sweep is $O(N)$. Data augmentation under $Z_2\\times D_4$ encodes the symmetries of the Hamiltonian into the model. The interpretation of the physics runs through the Allen-Cahn equation $v=-c\\kappa$: curvature-driven growth gives $t^{1/2}$, whereas straight walls and island domains at $T=0.01$ are consistent with a slower, corner-driven mechanism.","core_discovery":"The central claim is that a CNN with a fixed receptive field can act as an ML force field for discrete spins: given the Ising configuration in a $21\\times21$ block around site $i$, it outputs $\\Delta E_i = 2\\sigma_i h_i$, the energy change for flipping $\\sigma_i$. The model is trained on exact diagonalization of 100 snapshots of a $30\\times30$ lattice, augmented by $Z_2\\times D_4$ symmetry. Against held-out data the mean squared error is 0.0014, and the predicted critical temperature $T_c\\approx0.24$ matches prior KPM-based MCMC results. The paper's substantive physical discovery is that quenches of the square-lattice Ising double-exchange model show coarsening with a temperature-dependent exponent: $\\alpha\\approx1/2$ at $T=0.1$ (Allen-Cahn), $\\alpha\\approx1/4$ at $T=0.01$, and $\\alpha\\to0$ as $T\\to0$, with dynamical scaling holding at both temperatures. This is framed as unusual phase-ordering of a disorder-free itinerant system, attributed to frustrated long-range effective spin interactions or self-generated disorder from the spin-electron coupling.","pith_inferences":["If the $21\\times21$ locality cutoff is not sufficient at low temperatures, the reported $\\alpha=1/4$ could be an artifact of the surrogate; a direct convergence test against larger neighborhoods would settle this. This is our inference, not the paper's claim.","The temperature-dependent exponent resembles thermal-activated coarsening in disordered systems, so the Ising-DE model may provide a clean realization of self-generated disorder; comparing its scaling function to random-field or random-bond Ising coarsening would be a natural next step.","The ML framework is not limited to spin flips; any discrete degree of freedom coupled to itinerant electrons (e.g., Potts variables or occupational variables in phase-separating electron systems) could use the same locality-based training recipe.","A stronger test of the physics would be to compute the effective multi-spin couplings of Eq. (4) from the CNN and check whether the low-temperature long-range interactions are frustrated, as the paper suggests."],"forward_implications":["If the CNN surrogate is faithful, kinetic Monte Carlo of itinerant Ising systems becomes linear-scaling, so systems of $10^5$ spins and beyond are accessible without retraining.","The equilibrium benchmark pins $T_c\\approx0.24$ in agreement with prior KPM results, supporting that the ML energy model reproduces the thermodynamic universality class.","The coarsening result implies that the Ising double-exchange model has a temperature-dependent growth exponent $\\alpha(T)$, in contrast to short-range Ising models, and that $\\alpha$ vanishes in the zero-temperature limit.","Dynamical scaling in the form $C(r,t)=G(r/L(t))$ holds even when the growth law is anomalous, so the correlation function collapses onto a universal curve at each temperature.","The same architecture can be extended to q-state Potts or clock models by giving the network several output fields, one per possible local transition."],"supporting_citations":[{"why":"Supplies the Behler-Parrinello decomposition of energy into local atomic contributions that motivates the local-field approach.","marker":"[12]"},{"why":"Gives Kohn's linear-scaling argument that electronic structure is local, used to justify the finite-neighborhood assumption.","marker":"[27]"},{"why":"Formalizes the nearsightedness of electronic matter, the principle behind Eq. (6).","marker":"[28]"},{"why":"Provides prior KPM-based MCMC results for the Ising-DE transition temperature and 2D Ising universality used as the equilibrium benchmark.","marker":"[48]"},{"why":"Describes the kernel polynomial method, the linear-scaling electronic-structure baseline that still costs $O(N^2)$ per Monte Carlo sweep.","marker":"[52]"},{"why":"Defines the Glauber single-spin-flip dynamics used in the quench simulations.","marker":"[57]"},{"why":"Supplies the Allen-Cahn $t^{1/2}$ law and the dynamical-scaling framework that the paper tests against its measured exponents.","marker":"[58]"},{"why":"Gives the curvature-driven interface equation $v=-c\\kappa$ used to interpret the $T=0.1$ growth and to argue that straight walls freeze.","marker":"[61]"}],"fun_headline_variants":["ML force field speeds up Monte Carlo for itinerant magnets","CNN predicts spin flips to emulate magnetic dynamics","Slow coarsening in itinerant magnets linked to CNN surrogate","Machine learning unlocks large-scale simulations of Ising magnets","Temperature-dependent coarsening exposed by ML-driven Monte Carlo"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the assumption that the energy change of flipping a spin depends only on the spins inside a fixed $21\\times21$ block, and that this cutoff is accurate at all temperatures and domain shapes used; the paper does not show a convergence test against larger neighborhoods.","fun_headline_variants_meta":{"raw":{"variants":["ML force field speeds up Monte Carlo for itinerant magnets","CNN predicts spin flips to emulate magnetic dynamics","Slow coarsening in itinerant magnets linked to CNN surrogate","Machine learning unlocks large-scale simulations of Ising magnets","Temperature-dependent coarsening exposed by ML-driven Monte Carlo"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1717,"prompt_tokens":994,"completion_tokens":723,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":644}},"tokens_in":610,"tokens_out":723,"duration_ms":5740,"temperature":1.0,"reasoning_tokens":644,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:49:30.267104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact spin-flip energy by exact diagonalization for late-time $T=0.01$ snapshots of a $30\\times30$ system (straight walls and island domains) and compare with the CNN's prediction; if the typical error in $\\Delta E$ makes the acceptance probability $\\exp(-\\Delta E/T)$ deviate substantially from 1 at $T=0.01$, then the measured $\\alpha=1/4$ growth could be a surrogate artifact rather than a property of the model. A cleaner variant is to train a second network with a $31\\times31$ or $41\\times41$ neighborhood and check whether the fitted coarsening exponent changes.","supporting_citations":[{"cited_title":"Motome and N","cited_arxiv_id":null,"evidence_quote":"Provides prior KPM-based MCMC results for the Ising-DE transition temperature and 2D Ising universality used as the equilibrium benchmark."}],"review_version":1}