{"id":"2e51b142-8d02-429a-b77b-13aec6f65784","arxiv_id":"2412.11982","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A symmetry-aware echo state network trained on small-system charge-density-wave dynamics reproduces coarsening statistics on larger lattices and yields a growth exponent α≈0.375.","lead":"The paper trains a recurrent neural network, an echo state network, to predict how local charge-density-wave order evolves in a model electron-phonon system, then uses the trained network to simulate coarsening on lattices far larger than the training system. It reports that the machine-learned simulations reproduce the expected domain-growth statistics and find a growth exponent of about 0.375, below the standard Allen-Cahn value of 1/2.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The anomalous α=0.375 for CDW coarsening is measured only in closed-loop ESN simulations on a 120×120 lattice; with no exact baseline at that scale and no neighborhood-size convergence test, the exponent could be a surrogate artifact.","rationale":"The TDGL benchmark (Sec. III) shows the ESN architecture can reproduce known Allen-Cahn coarsening, and the 40×40 CDW correlation comparison gives nontrivial support that the local ESN is not trivially wrong. Those checks, however, are at scales and times where the exact simulation is available; the interesting result, α=0.375 at 120×120, is precisely where no exact comparison is made. Because the surrogate is iterated in a closed loop, systematic errors in the learned local map can accumulate into a biased growth law. This is the single most load-bearing assumption in the paper: the 7×7, φ-only closure of Eq. (5) must hold well enough at 120×120 that the statistical coarsening is unchanged. The proposed KPM computation would settle whether the extracted exponent is physical or an artifact. Since the reader's CONDITIONAL verdict already requires exactly this kind of additional validation, the verdict should remain UNCHANGED.","tokens_in":44944,"tokens_out":7263,"duration_ms":74201,"concrete_test":"Run a KPM-based Langevin simulation (the linear-scaling fermion solver cited in Sec. IV.A) on a 120×120 lattice with the same parameters (g=1, k=1, κ=0.18, T=0.001, ωτ∼0.05, δt=0.0025 t0) and at least 30 independent quenches. Compute L(t) via Eq. (15) over the same nstep range as the ESN runs and fit α with confidence intervals. If the exact α matches 0.375, the ESN result is supported; if it is closer to the Allen-Cahn value 1/2 or differs beyond the fit errors, the headline exponent is an ESN artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central physics claim depends on Eq. (5): that the local CDW order-parameter field closes under its own dynamics in a 7×7 window. The true microstate is (Q_i, P_i), with P_i generated by the Langevin equation (13), and φ_i is a derived electron-density observable (Eq. (12)); the authors themselves note in Sec. IV.B that electron-mediated interactions are longer-ranged and that it is unclear whether φ obeys a closed PDE. The single-site clamped benchmark in Fig. 5 and the 40×40 correlation check in Fig. 6 are encouraging, but they do not validate the closed-loop 120×120 trajectory from which α=0.375 is extracted. No error bars are reported for the exponent, and its only independent-looking support is a same-group ML force-field paper [46]. This is a validation gap, not an internal inconsistency: if the ESN's learned local map and reservoir memory happen to distort long-time coarsening statistics, the headline anomalous exponent would not represent Holstein physics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a local echo-state-network (ESN) surrogate for the phase-ordering dynamics of a scalar order parameter, with emphasis on the charge-density-wave (CDW) order of the square-lattice Holstein model. After benchmarking the ESN on the time-dependent Ginzburg-Landau (TDGL) equation, where it recovers the Allen-Cahn L(t)∼t^{1/2} collapse, the authors train a symmetry-constrained ESN on one 40×40 ED-Langevin quench, validate bare correlation functions at three times on the same size, and then iterate the ESN on 120×120 lattices. From the large-scale surrogate they report dynamical scaling and a growth exponent α=0.375, which they interpret as anomalous coarsening caused by longer-ranged, electron-mediated interactions.","tokens_in":45171,"tokens_out":5315,"duration_ms":47685,"significance":"If the 0.375 exponent is robust, this is a valuable result: it demonstrates that a local reservoir-computing surrogate can transfer from small to large lattices and offers a large speedup over ED-Langevin simulations for coarsening statistics, opening a route to large-scale modeling of electron-lattice phase ordering. The TDGL benchmark is a genuine success: the ESN reproduces both the scaling function and the expected t^{1/2} law using a training set from a 30×30 system, and the D4-symmetry-aware input construction is a sensible architectural choice. However, the central CDW claim rests on a single closed-loop ESN simulation at 120×120 without a direct exact baseline, without error bars, and without a neighborhood-size convergence test, so the paper's main quantitative conclusion is not yet established.","major_comments":[{"comment":"The exponent α=0.375 is extracted solely from ESN-generated correlation lengths on a 120×120 lattice, but the manuscript reports no exact ED-Langevin (or KPM) simulation at that scale, no independent realizations, and no error bar on the fitted slope. Since the authors themselves note in §IV.B that closed-loop ESN trajectories quickly deviate from exact dynamics, the extracted exponent could be a property of the surrogate rather than of the Holstein model. Please provide a direct comparison at a size accessible to an exact or KPM baseline, statistics over multiple ESN trainings and quenches, and an explicit statement of the fitting range used for the log-log slope.","section":"§IV.C, Fig. 8 inset"},{"comment":"The framework assumes ϕ_i(t+Δt)=F({ϕ_j(t)|r_j∈N_i}) with a 7×7 window, but the paper acknowledges in the same section that the electron-mediated interaction is longer-ranged and that it is unclear whether ϕ obeys a closed PDE. The choice ℓ=7 is not tested: no results with different neighborhood sizes (e.g., 5×5, 7×7, 9×9) are shown, and no comparison of the resulting L(t) exponent as a function of ℓ is reported. Without such a convergence test, the locality assumption, which is the basis for transferability and for the 120×120 claim, is unsupported.","section":"§IV.B, Eq. (5)"},{"comment":"The 40×40 validation reports bare correlation functions at three times and states that they agree very well with ED-Langevin data, but it does not report the corresponding L(t) growth law from the exact ED-Langevin simulations, nor a quantitative error measure on the correlations. Agreement of correlation functions at three early-to-intermediate times on the training-size lattice is necessary but not sufficient to show that the ESN reproduces the coarsening exponent, because that exponent is determined by the late-time evolution of L(t), which is exactly where the ESN predictions are known to deviate from exact trajectories.","section":"§IV.B, Fig. 6"},{"comment":"The only external-looking support for α=0.375 is Ref. [46], a prior machine-learning force-field study from the same group. This cannot serve as an independent validation of the surrogate result. Please compare with a non-ML method (exact ED-Langevin or KPM at moderate sizes) or with independent literature values, and if none are available, weaken the claim accordingly.","section":"§IV.C, Ref. [46]"}],"minor_comments":[{"comment":"There are typographical errors that should be corrected, including 'every sites one the lattice' in §II.B and 'scalable ESN-based scalable ML approach' in §IV.B.","section":"§II.B and §IV.B"},{"comment":"The statement 'nearly perfect data-point collapsing' is not supported by a quantitative collapse measure; consider reporting a residual or an overlap metric for the scaling collapse.","section":"§IV.C, Fig. 8"},{"comment":"The fixed hyperparameters (reservoir size, spectral radius, connectivity, ridge regularization, input scaling, and prediction time step) are chosen without a sensitivity analysis; even a brief robustness check would be helpful, especially because the central exponent is extracted from a single parameter choice.","section":"§IV.B, hyperparameters"},{"comment":"The correlation-length definition in Eq. (15) is applied on a 120×120 periodic system, but the authors note in §IV.B that C(r,t) does not decay to zero for r at half the linear size; the sums in Eq. (15) may therefore be sensitive to finite-size offsets. Please state the largest r included in the sums or subtract the long-distance plateau.","section":"§IV.C, Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision. The paper is within scope and the TDGL benchmark is strong, but the central CDW exponent needs a direct baseline, uncertainty quantification, and a locality-convergence test. The reliance on Ref. [46] from the same group as the only cross-check is a concern for the editor, though not by itself a reason for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my take. The real contribution is the method: a local echo state network with D4-symmetrized distance-shared input weights, trained on a single small ED-Langevin quench, then transferred to a 120x120 lattice. The TDGL benchmark is solid; the ESN reproduces the Allen-Cahn t^{1/2} collapse, which tells you the architecture and training are sound. The transfer from 40x40 to 120x120 is the new piece, and the correlation-function match at 40x40 at three times is a fair check. The claimed speedups (10^4 measured, 10^7 extrapolated) are clearly stated and not absurd. The soft spot is exactly where the stress-test note lands. The central physics result, alpha=0.375, comes only from the closed-loop ESN on 120x120. There is no ED-Langevin run at that scale, no error bars, and no test of whether a larger neighborhood changes the answer. The authors say themselves that electron-mediated effective interactions are longer-ranged than in the TDGL case and that it is not clear whether phi has a closed PDE. The single-site clamped benchmark and the 40x40 correlation check do not validate the long-time statistics at 120x120. So the exponent could be a surrogate artifact. That is a validation gap, not an internal contradiction. One more thing: the only independent-looking support for alpha=0.375 is a previous ML force-field paper from the same group. That is not circular, since the exponent is not a fitted parameter, but it is far from external confirmation. No code or data are released, so the numbers cannot be re-run. Bottom line: this deserves a serious referee. The method is a real step forward and the physics question is worth asking, but the headline claim needs more support before it should be published as a result. A careful referee should ask for error bars on the exponent, a neighborhood-size convergence study, and at least one exact or KPM-based baseline at the larger scale. I would send it to review, not desk-reject it, and push for those additions.","headline":"The method is genuinely useful, but the headline alpha=0.375 rests on the surrogate alone and needs an exact baseline before it earns the status of a physics result.","tokens_in":45733,"tokens_out":2771,"would_cite":true,"duration_ms":25373,"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":"Training a local echo state network on one small Holstein-model quench produces a transferable simulator whose CDW domain growth follows $L(t)\\sim t^{0.375}$, slower than the Allen-Cahn law.","keywords":["echo state network","reservoir computing","charge density wave","Holstein model","phase ordering kinetics","domain growth","Allen-Cahn law","dynamical scaling"],"falsifier":"Run a direct ED-Langevin simulation of the same Holstein model on the largest lattice that is still tractable (for example, $60\\times60$ or $80\\times80$) with identical parameters, extract the domain-growth exponent from the correlation length, and compare it with the ESN's $\\alpha=0.375$; if the exact dynamics gives $\\alpha$ close to $1/2$ or outside the ESN's error bars, the slower exponent is a surrogate artifact. A cheaper complementary test is to retrain the ESN with $9\\times9$ and $11\\times11$ neighborhoods and check whether the predicted $\\alpha$ remains stable as the window grows.","tokens_in":44708,"feed_emoji":"⚡","tokens_out":15784,"duration_ms":126491,"temperature":0.7,"pith_summary":"The paper sets out to show that a local echo state network—a recurrent neural network trained only through a linear readout—can learn the emergent coarsening dynamics of charge-density-wave (CDW) order in the semi-classical Holstein model and then be reused as a linear-scaling simulator on lattices much larger than the training system. It first benchmarks the approach on the time-dependent Ginzburg-Landau equation, where the ESN recovers the Allen-Cahn growth law $L(t)\\sim t^{1/2}$. It then trains the ESN on a single exact-diagonalization Langevin quench of a $40\\times40$ Holstein lattice, with local CDW order parameters in a $7\\times7$ neighborhood as input, and iterates it on a $120\\times120$ lattice. The central result is that the ESN reproduces the statistical coarsening of CDW domains and yields a domain-growth exponent $\\alpha=0.375$, slower than the Allen-Cahn value, which the paper attributes to longer-ranged electron-mediated interactions. A sympathetic reader would care because this is a concrete route to simulating phase-ordering kinetics in complex electron systems where the driving forces are too expensive to compute exactly every time step.","feed_headline":"Trained on 40x40, echo-state net simulates CDW growth at 120x120","feed_subtitle":"It reproduces the statistics of coarsening and yields an exponent of 0.375, below the Allen-Cahn 1/2.","key_machinery":"The central object is a local echo state network, defined in one phrase as a recurrent network with a fixed random reservoir and a linear readout trained by ridge regression. The network implements the update map $\\phi_i(t+\\Delta t)=F(\\{\\phi_j(t)\\mid r_j\\in\\mathcal{N}_i\\})$ of Eq. (5), with $\\mathcal{N}_i$ chosen as a $7\\times7$ block for the CDW problem; neighbors farther away are assumed irrelevant or already encoded in the reservoir state. The symmetry requirement enters through Eq. (14), where the input weight from any neighbor to a reservoir neuron depends only on that neighbor's distance from the center, plus a $1/r$ influence decay, so the map is covariant under the rotations and reflections of the square lattice. The reservoir's spectral radius $\\rho=0.79$ and the recurrent connections provide the memory needed for dynamics that is not a simple first-order Markov map. This machinery carries the claim of transferability: because the input tensor size is fixed by the neighborhood, the trained network can be swept over any lattice size at linear cost.","core_discovery":"The paper argues that the effective dynamics of the CDW order parameter in the Holstein model closes locally: the value $\\phi_i(t+\\Delta t)$ at a site depends on the configuration of $\\phi$ in a finite neighborhood $\\mathcal{N}_i$ at time $t$, supplemented by the reservoir's memory of the recent past. Using Eq. (5) as the design principle, the authors build an ESN whose input weights respect the $D_4$ point group of the square lattice (neighbors at equal distance share coupling), train it on 400 configurations from one ED-Langevin quench of a $40\\times40$ system at low temperature, and then apply the same network site-by-site to simulate a $120\\times120$ lattice. For the TDGL benchmark the ESN correctly produces the Allen-Cahn law $\\alpha=1/2$ and the associated scaling collapse, even though individual trajectories diverge from exact dynamics. For the Holstein CDW, the ESN's correlation functions agree with exact ED-Langevin results on the $40\\times40$ lattice, and the large-scale simulation shows clean dynamical scaling with $L(t)\\sim t^\\alpha$, $\\alpha=0.375$, a value the paper interprets as a genuine consequence of electron-mediated longer-ranged and frustrated interactions rather than a neural-network artifact. The paper states that this value is consistent with an earlier machine-learning force-field study of the same model.","pith_inferences":["The paper does not check convergence of the growth exponent as the neighborhood grows beyond $7\\times7$; an editorial inference is that $\\alpha$ should be tested with $9\\times9$ and $11\\times11$ inputs, and if it shifts, the surrogate is not yet faithful to the true interaction range.","If the slower exponent is physical, the coarsening of CDW order likely belongs to a different dynamical universality class than the Allen-Cahn one; a natural extension is to measure two-time quantities such as the autocorrelation exponent and compare with the values expected for long-range or frustrated coarsening.","The same local-ESN construction could be applied to other emergent order parameters without a known closed equation, such as spin-density waves or superconducting phase patterns, by training on data from a single small simulation.","A testable extension is to feed time-resolved experimental images of CDW domains into the same architecture and compare the extracted growth exponent with the simulation value, which would directly probe whether the Holstein model's coarsening physics is realized in materials."],"forward_implications":["A single ESN trained on a $40\\times40$ Holstein quench can be applied directly to a $120\\times120$ lattice, giving linear-scaling simulations of CDW coarsening without retraining.","The ESN captures statistical coarsening (correlation functions, scaling collapse, growth exponent) even when pointwise trajectories diverge, so it can serve as a surrogate for 'climate' quantities in phase ordering.","The predicted CDW exponent $\\alpha=0.375$ is slower than the Allen-Cahn $1/2$, implying that electron-mediated interactions modify the growth law expected from short-ranged Ising symmetry.","Because the ESN output step $\\Delta t=0.05\\,t_0$ is many times larger than the Langevin integration step, the surrogate is roughly $10^4$ times faster than ED-Langevin at equal size and projected to be about $10^7$ times faster at $120\\times120$.","The $D_4$-symmetrized input weights are needed for accuracy; an unconstrained random input matrix degrades the prediction, which the paper demonstrates directly."],"supporting_citations":[{"why":"Defines the echo state network architecture, echo state property, and ridge-regression training that the paper's surrogate is built on.","marker":"[9]"},{"why":"Provides the dynamical-scaling framework and Allen-Cahn growth law against which the ESN's TDGL and CDW results are judged.","marker":"[25]"},{"why":"Earlier machine-learning force-field study of Holstein CDW phase ordering whose growth exponent the paper says is consistent with the ESN's alpha = 0.375.","marker":"[46]"},{"why":"Previous ESN prediction of phase-ordering dynamics that showed short-term accuracy with long-term statistical capture, the pattern the paper benchmarks.","marker":"[48]"},{"why":"The Allen-Cahn analysis that gives $L(t)\\sim t^{1/2}$ for curvature-driven growth of a non-conserved Ising order parameter, the baseline for the anomalous CDW exponent.","marker":"[52]"},{"why":"Introduces the Holstein model whose semi-classical electron-lattice dynamics generates the CDW coarsening being learned.","marker":"[53]"},{"why":"The nearsightedness/locality principle that justifies the fixed finite neighborhood and the transfer of the network to larger lattices.","marker":"[24]"}],"fun_headline_variants":["Echo-state net scales CDW coarsening from 40x40 to 120x120","Small-trained ESN simulates large-scale charge density waves","CDW coarsening exponent 0.375 from echo-state network","ESN reveals CDW growth below Allen-Cahn limit","Neural reservoir predicts CDW coarsening across scales"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the future CDW order at a site is determined by the current order parameters in a fixed $7\\times7$ block around it, so anything happening farther away either does not matter or is already captured in that local picture.","fun_headline_variants_meta":{"raw":{"variants":["Echo-state net scales CDW coarsening from 40x40 to 120x120","Small-trained ESN simulates large-scale charge density waves","CDW coarsening exponent 0.375 from echo-state network","ESN reveals CDW growth below Allen-Cahn limit","Neural reservoir predicts CDW coarsening across scales"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1503,"prompt_tokens":1084,"completion_tokens":419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":700,"completion_tokens_details":{"reasoning_tokens":326}},"tokens_in":700,"tokens_out":419,"duration_ms":4701,"temperature":1.0,"reasoning_tokens":326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:24:22.142266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a direct ED-Langevin simulation of the same Holstein model on the largest lattice that is still tractable (for example, $60\\times60$ or $80\\times80$) with identical parameters, extract the domain-growth exponent from the correlation length, and compare it with the ESN's $\\alpha=0.375$; if the exact dynamics gives $\\alpha$ close to $1/2$ or outside the ESN's error bars, the slower exponent is a surrogate artifact. A cheaper complementary test is to retrain the ESN with $9\\times9$ and $11\\times11$ neighborhoods and check whether the predicted $\\alpha$ remains stable as the window grows.","supporting_citations":[{"cited_title":"echo state","cited_arxiv_id":null,"evidence_quote":"Defines the echo state network architecture, echo state property, and ridge-regression training that the paper's surrogate is built on."},{"cited_title":"Chauhan, S","cited_arxiv_id":null,"evidence_quote":"Previous ESN prediction of phase-ordering dynamics that showed short-term accuracy with long-term statistical capture, the pattern the paper benchmarks."},{"cited_title":"Allen and J","cited_arxiv_id":null,"evidence_quote":"The Allen-Cahn analysis that gives $L(t)\\sim t^{1/2}$ for curvature-driven growth of a non-conserved Ising order parameter, the baseline for the anomalous CDW exponent."},{"cited_title":"Holstein, Studies of polaron motion: Part i","cited_arxiv_id":null,"evidence_quote":"Introduces the Holstein model whose semi-classical electron-lattice dynamics generates the CDW coarsening being learned."}],"review_version":1}