{"id":"29be50aa-198a-4928-b9c4-31041f381e72","arxiv_id":"1908.04604","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A graphene strain superlattice made of repeated Gaussian bumps produces valley-polarized conductance plateaus, and a deep neural network can approximate the valley polarization computed by Green's functions.","lead":"The paper shows that a periodic chain of Gaussian bumps on graphene acts as a valley filter, transmitting electrons from one valley while blocking the other, in wider energy ranges than a single bump. The authors also train a deep neural network to predict this valley polarization from design parameters, aiming for faster searches over possible superlattices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random-row train/test split likely inflates the DNN generalization claim: the reported R2=0.970 may reflect energy interpolation within geometries seen in training, not prediction for unseen superlattices; a configuration-level holdout test is needed.","rationale":"The Green's function physics part is presented with a standard mode-matching method, and the identification of mini-stopbands through Eq. 6 is internally consistent with the band-structure plots and plateau positions. The disorder robustness study is a useful internal consistency check. The weakest link is the machine learning generalization claim: the random-row split means each configuration's 500 energy points are distributed across training and test, so neighboring-energy interpolation can produce a high R2 even if the network cannot predict for an unseen superlattice geometry. The paper's own admission that the DNN does not reproduce plateau widths is especially damaging because the design metric QK' quantifies exactly those widths. The single Green's function validation of the DNN-selected optimum (Fig. 9b) does not resolve the issue unless that geometry is shown to have been excluded from training. A grouped-holdout evaluation is the decisive check. Until that is reported, the reader's CONDITIONAL verdict is appropriate and unchanged.","tokens_in":15994,"tokens_out":8368,"duration_ms":94443,"concrete_test":"Retrain the same DNN architecture with configuration-level grouped cross-validation: partition the 117 superlattice configurations (not the 58,500 rows) into folds so that all 500 energy samples of each geometry stay together in either training or test, and compute held-out R2 and MSE on entire unseen configurations. If the configuration-level R2 falls materially below 0.970, the random-row split is the cause and the Section III B design-search claims are not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The DNN half of the central claim rests on the generalization metric reported in Section III A. The 58,500 rows come from only 117 superlattice configurations, each contributing 500 energy points, and the paper says the data were 'randomly split' by rows into 80% training and 20% testing. This permits the same geometry to appear in both training and test sets, so the model can interpolate in energy for geometries already seen during training. The reported R2=0.970 therefore measures energy interpolation, not the ability to predict valley polarization for a new superlattice geometry. That distinction is load-bearing because the design-tool value claimed in Section III B depends precisely on generalization to unseen configurations. The paper itself concedes that the DNN fails to reproduce plateau widths and low-polarization oscillations, and QK' is a width-based figure of merit, so the design-search conclusions are not protected by the high R2. The one Green's function check of the DNN-selected optimum (Fig. 9b) would only be probative if that configuration was not part of the training set, which the paper does not establish.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies valley-polarized electron transport through a zigzag graphene nanoribbon containing a one-dimensional superlattice of out-of-plane Gaussian deformations. Using a lattice Green's function / wave-function-matching method, the authors show that periodicity produces fully valley-polarized conductance plateaus (PK' = 1) and attribute this to the coupling of counter-propagating transverse modes that opens mini-stopbands, with Bragg energies given by Eq. (6). They also train a deep neural network on data generated by the same Green's-function method to predict valley polarization as a function of six design parameters, report R2 = 0.970 on a test set, and use the DNN to search for an optimal superlattice, which they identify as NG = 6 and α = 15%. The DNN predictions are evaluated against the same Green's-function technique that generated the training data.","tokens_in":16187,"tokens_out":5536,"duration_ms":53300,"significance":"The Green's-function part of the paper is careful and the proposed mechanism—mini-stopbands opened by periodicity-induced mode coupling—is supported by the band-structure analysis and by the quantitative agreement between the Bragg energies of Eq. (6) and the positions of the valley-filter plateaus. The demonstration that a periodic array of Gaussian bumps yields valley filtering at lower strain and in wider energy windows than a single bump is a valuable result for the valleytronics literature. The DNN surrogate is an interesting and potentially useful tool, but the reported generalization metric is compromised by the data-splitting procedure. Because the design-tool claim in Section III B rests on generalization to unseen superlattice configurations, the machine-learning part of the paper requires re-validation. If a configuration-level holdout test confirms good accuracy, the paper will be a solid contribution; as it stands, the machine-learning claims are overstated.","major_comments":[{"comment":"The reported test accuracy R2 = 0.970 is obtained by randomly splitting the 58,500 energy samples into 80% training and 20% testing. Because each of the 117 configurations contributes 500 energy points, this row-wise split places energies from the same superlattice geometry in both the training and test sets. The reported R2 therefore measures interpolation in energy for geometries already seen during training, not prediction for a new (Nx, Ny, NG, b, d, α) configuration. The design-tool claims in §III B require the latter. Please repeat the evaluation with a configuration-level split (for example, hold out whole configurations or use leave-one-configuration-out cross-validation) and report the resulting R2 and MSE. It would also be informative to state whether the selected optimum (NG = 6, α = 15%) was present in the training set under the original split.","section":"§III A"},{"comment":"The text contains a direct internal contradiction about the DNN's accuracy on plateau width. It states that \"The DNN accurately predicts the energy location of the valley plate but fails to estimate its width\" and, in the discussion of Fig. 9a, that DNN-predicted QK' values are \"smaller than the true values calculated by the Green's functions.\" Yet the caption of Fig. 9b says the DNN predicts \"all the energy position and width of the valley plateaus with acceptable accuracy.\" Since QK' is a width-based figure of merit (Eq. (7)), a systematic underestimation of plateau width means the design-search result is not protected by the high R2. Please quantify the DNN error in QK' on configurations not used in training and reconcile the two statements.","section":"§III B"},{"comment":"The design-search paragraph does not specify the objective used to trade off QK', α, and NG. The claim that NG = 6 and α = 15% is \"optimal\" depends on an unstated ranking criterion. Please define the criterion (for example, a scalarized objective or a Pareto-front definition) so that the search is reproducible and the result is interpretable.","section":"§III B"}],"minor_comments":[{"comment":"The term \"valley notch filter\" is used throughout, but the described effect is a fully valley-polarized passband (a plateau where PK' = 1), not a narrow notch of rejection. Please clarify the terminology or justify the use of \"notch filter\" in the abstract and introduction.","section":"Abstract / Title"},{"comment":"The phrase \"a sequence by valley filter plateaus\" is ungrammatical; it should read \"a sequence of valley filter plateaus.\" Similar grammar issues appear elsewhere (for example, \"the firsts valley filter plateaus\" in §II B).","section":"Abstract"},{"comment":"The caption of Fig. 2 lists transmission probabilities for panels (d) and (e), but the body text refers to Fig. 2(e)-(f) for the same quantities. Please make the panel references consistent.","section":"Fig. 2 caption"},{"comment":"The dataset summary includes configurations with NG = 0, b = 0, d = 0, and α = 0. The physical meaning of these boundary cases (presumably pristine ribbons) should be stated explicitly, as they may affect the DNN training and the reported statistics.","section":"Table II"},{"comment":"The phrase \"hyperparameter tunning\" should be \"hyperparameter tuning\".","section":"§III A"},{"comment":"The word \"supperlattice\" is misspelled as \"superlattice\" in several places, including §II B and the caption of Fig. 5.","section":"Various"},{"comment":"The computational-complexity statement \"inverting a matrix is O(Ny^γ) where 2.3 ≤ γ ≤ 3\" is vague; please specify which matrix is inverted and which algorithm the bound refers to.","section":"§IV"}],"recommendation":"major_revision","confidential_remarks":"The main technical concern is the train/test split in the machine-learning section. This is a fixable issue, but it requires re-running the experiments with a configuration-level holdout and revising the design-tool claims accordingly. The Green's-function physics appears sound, and the manuscript should be publishable after the ML validation is corrected. I would also like to see the design-search ranking criterion specified. No concerns about scope or attribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Couple of things you should know before reading. The physics claim is the real contribution: a periodic chain of Gaussian bumps turns a single-bubble valley filter into a robust notch filter with fully polarized plateaus, and the authors identify the mechanism as counter-propagating mode coupling opening mini-stopbands. The band-structure and Green's function calculations agree, and the strain required is lower than for single bubbles. That part is solid, and it is new relative to the cited nanobubble work. The machine-learning part is the weak half. The DNN is trained on 58,500 energy samples from just 117 superlattice geometries, and the split is random by rows, not by configuration. So the 20% test set contains energies from the same geometries seen in training; R2=0.970 mostly measures energy interpolation for known geometries, not generalization to a new superlattice. The paper itself concedes the DNN does not reproduce plateau widths and low-polarization oscillations, and QK' is a width-based metric, so the design-search conclusions in Section III B are not protected by the high R2. The Green's function check of the DNN-selected optimum (Fig. 9b) is only probative if that configuration was not in training, which is not established. This is a real flaw, but it does not sink the physics: the mechanism and plateaus are computed independently of the DNN. What is missing is a configuration-level holdout, ideally with code and data released, before the ML claims are accepted. Minor issues: the disorder study uses only 20 realisations at one disorder level, which is fine but limited; the statement that wrinkles or gate-potential variations will not change the results is an expectation, not a calculation. The citation pattern is fine, covering the relevant nanobubble, superlattice, and ML literature. Bottom line: the valley notch filter result deserves refereeing and likely publication after the ML validation is redone. The DNN section needs work; the physics does not.","headline":"Physics of the valley notch filter is solid and novel; the DNN generalization claim is inflated by a random-row split, but that is fixable and the paper deserves refereeing.","tokens_in":16737,"tokens_out":1828,"would_cite":true,"duration_ms":19064,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["73.23.-b","73.63.-b","81.05.ue","07.05.Mh"],"model":"deepseek-v4-flash","headline":"Periodic Gaussian strain bumps turn a graphene nanoribbon into a valley notch filter that fully polarizes one valley.","keywords":["valleytronics","graphene nanoribbon","strain superlattice","pseudo-magnetic field","valley filter","Green's function","deep neural network","mini-stopbands"],"falsifier":"Train the same network on a split that keeps all 500 energy samples of each of the 117 device layouts together in either the training or the test set, and measure $R^2$ on the held-out layouts; a large drop from the reported 0.970 would show the network has not generalized to unseen superlattices. On the physics side, the stated mechanism would be contradicted if a band-structure calculation of the infinite Gaussian chain showed no mini-stopband anticrossing at the Bragg energies of Eq. (6).","tokens_in":15768,"feed_emoji":"🧲","tokens_out":14875,"duration_ms":139286,"temperature":0.7,"pith_summary":"This paper tries to establish that arranging Gaussian out-of-plane bumps into a periodic chain turns a strained graphene ribbon into a much stronger valley filter than a single bump. The periodicity folds the ribbon's transverse-mode bands and couples modes travelling in opposite directions; at the Bragg energies of Eq. (6) this opens mini-stopbands that reject one valley, producing fully valley-polarized conductance plateaus over wider energy ranges and at lower strain than single-bubble devices. The paper also claims that the same physics is learnable: a deep neural network trained on Green's-function results predicts the valley polarization almost as accurately as the full calculation while costing far less computation, and a design search with the network identifies a six-bump, low-strain chain as an optimal valley filter. If these claims hold, valleytronic devices would not need perfect zigzag edges, high strain, or carefully tuned energies.","feed_headline":"Strain bumps turn graphene into a valley notch filter","feed_subtitle":"A periodic chain of out-of-plane bumps fully polarizes one valley at low strain, and a neural network predicts the effect fast.","key_machinery":"The load-bearing object is the Bragg condition for counter-propagating transverse modes, Eq. (6): $$E^l_{mm} = (3 t_0 a_c/2)\\sqrt{(l\\pi/d)^2 + (\\pi(m+1/2)/W_0)^2}.$$ It fixes the energies at which mode $m$ is reflected by the superlattice period $d$; the resulting anticrossings open mini-stopbands, and those stopbands are the notch filter. The transport calculation uses the lattice Green's function with wave-function matching to resolve transmission by valley, and the machine-learning component is a seven-hidden-layer feedforward network that maps seven design inputs to the valley polarization.","core_discovery":"This paper's central claim is that a one-dimensional chain of out-of-plane Gaussian deformations in a zigzag graphene nanoribbon acts as a valley notch filter: the conductance develops quantized plateaus with full valley polarization in the $K'$ valley ($P_{K'}=1$) while the $K$ valley is blocked. The mechanism is the periodic folding of the transverse-mode spectrum combined with strain-induced pseudo-magnetic fields; counter-propagating modes in the same valley couple at the Bragg energies of Eq. (6), opening mini-stopbands that selectively reject one valley. The paper argues this is a genuine improvement over a single Gaussian bump, where valley imbalance comes only from the extra edge-state mode, requires fine-tuned energy, and demands high strain; the superlattice produces the filter from low-energy bulk modes, works at lower strain and over wider energy windows, and is robust to moderate disorder in bump height, width, and spacing. The paper further claims that a deep feedforward network trained on Green's-function data predicts $P_{K'}$ with reported $R^2=0.970$ and mean squared error $0.001$ on the test split, at much lower computational cost, and uses the network to propose a six-bump, $\\alpha=15\\%$ chain as the best low-strain compact valley filter in the scanned design space.","pith_inferences":["Inference: Because the Bragg condition in Eq. (6) depends only on period, ribbon width, and transverse-mode index, not on the Gaussian profile, the same notch-filter mechanism should appear for other periodic out-of-plane profiles, such as sinusoidal ripples or folded geometries; a Green's-function calculation for those profiles would test that.","Inference: The reported tendency of the neural network to locate plateaus but underestimate their widths suggests the design-tool value would improve with a loss function that weights the plateau edges more heavily, or with a two-stage model that first classifies plateau regions and then regresses their widths.","Inference: The DNN result that a few bumps mimic an infinite chain implies the practical device could be short, and an experimental signature to look for is the emergence of quantized plateaus as the number of nanopillars is increased from one to about six."],"forward_implications":["A periodic Gaussian chain fully valley-polarizes the conductance in quantized plateaus ($P_{K'}=1$), in contrast to a single bump's weak polarization confined to the edge-state mode.","The plateau energies are controlled by the Bragg condition of Eq. (6), so the filter's operating windows can be positioned by choosing the bump spacing $d$ and ribbon width $W_0$.","Higher-energy plateaus are carried by low-energy bulk transverse modes, making the filter less vulnerable to edge roughness than single-bubble devices.","The filter remains effective under roughly 10% disorder in bump height, width, and spacing; the reported valley-filter capability changes from 22.5% to between 20.7% and 22.7% for the tested disorder.","The trained DNN reproduces $P_{K'}$ with $R^2=0.970$ on the reported test split at much lower cost than the Green's-function solve, and the resulting design search identifies a compact $N_G=6$, $\\alpha=15\\%$ chain as optimal among the scanned parameters."],"supporting_citations":[{"why":"It supplies the wave-function matching and Green's function method used to compute valley-resolved transmission probabilities.","marker":"[40]"},{"why":"It provides the mini-stopband mechanism for transverse-mode coupling in graphene nanoribbons that the superlattice filter adapts.","marker":"[28]"},{"why":"It demonstrates single-nanobubble valley filtering, the baseline the paper's superlattice is claimed to improve on.","marker":"[16]"},{"why":"It establishes that strained graphene nanobubbles generate pseudo-magnetic fields, the physical basis for the valley-dependent gauge field.","marker":"[15]"},{"why":"It reports experimental strain superlattices with pseudo-Landau levels, supporting the practical feasibility of the proposed device geometry.","marker":"[29]"},{"why":"It gives the gauge-field description of strain in graphene used to define the pseudo-vector potential in Eq. (3).","marker":"[38]"}],"fun_headline_variants":["Periodic strain bumps form graphene valley notch filter","Neural network speeds up graphene valley filter prediction","Graphene strain superlattice yields complete valley polarization","Periodic deformations turn graphene into valley notch filter","Deep learning predicts graphene valley filter from strain bumps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the network's reported test accuracy, measured on energy values drawn from the same device layouts that were used in training, also holds for device layouts it has never seen; if that generalization fails, the fast design-tool claim collapses even though the Green's-function physics may stand.","fun_headline_variants_meta":{"raw":{"variants":["Periodic strain bumps form graphene valley notch filter","Neural network speeds up graphene valley filter prediction","Graphene strain superlattice yields complete valley polarization","Periodic deformations turn graphene into valley notch filter","Deep learning predicts graphene valley filter from strain bumps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2513,"prompt_tokens":951,"completion_tokens":1562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":1489}},"tokens_in":567,"tokens_out":1562,"duration_ms":16745,"temperature":1.0,"reasoning_tokens":1489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:37:16.759339+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same network on a split that keeps all 500 energy samples of each of the 117 device layouts together in either the training or the test set, and measure $R^2$ on the held-out layouts; a large drop from the reported 0.970 would show the network has not generalized to unseen superlattices. On the physics side, the stated mechanism would be contradicted if a band-structure calculation of the infinite Gaussian chain showed no mini-stopband anticrossing at the Bragg energies of Eq. (6).","supporting_citations":[{"cited_title":"Khomyakov , author G","cited_arxiv_id":null,"evidence_quote":"It supplies the wave-function matching and Green's function method used to compute valley-resolved transmission probabilities."},{"cited_title":"Benisty ,\\ title title Graphene nanoribbons: Photonic crystal waveguide analogy and minigap stripes ,\\ https://doi.org/10.1103/PhysRevB.79.155409 journal journal Phys","cited_arxiv_id":null,"evidence_quote":"It provides the mini-stopband mechanism for transverse-mode coupling in graphene nanoribbons that the superlattice filter adapts."},{"cited_title":"Jiang , author J","cited_arxiv_id":null,"evidence_quote":"It reports experimental strain superlattices with pseudo-Landau levels, supporting the practical feasibility of the proposed device geometry."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It gives the gauge-field description of strain in graphene used to define the pseudo-vector potential in Eq. (3)."}],"review_version":1}