{"id":"5ad81705-cc01-4f02-a777-fc31390332d6","arxiv_id":"2412.11129","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A CNN trained on model-generated ARPES spectra is used to claim that bilayer splitting in underdoped BSCCO does not decrease with underdoping, though the evidence is model-dependent and limited to two samples.","lead":"This paper trains a convolutional neural network on synthetic photoemission spectra to detect bilayer splitting in cuprate superconductors, and applies the network to two experimental spectra. It claims the splitting persists in underdoped BSCCO, but the analysis depends on a generative model that already assumes the splitting it aims to confirm.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that underdoping does not reduce bilayer splitting rests on untested transfer of a CNN from a single generative model to real ARPES spectra; without a negative control, the network may be detecting peak shape rather than two bands.","rationale":"The reader's weakest assumption identifies the generative model's faithfulness as the key risk, and this stress test agrees: the most load-bearing link in the paper is the implicit claim that a CNN trained on Eqs. (1)-(8) transfers to real ARPES spectra without any independent validation. I sharpen this into a concrete failure mode: the trained network may exploit the presence of a single broad peak with strong self-energy broadening as evidence of splitting, because the training set never contains single-band spectra with the large resonance self-energy that characterizes the underdoped superconducting state. The paper's strongest physical conclusion, that underdoping does not reduce bilayer splitting, is drawn from exactly two real spectra, both of which lie in this untested regime. This is not a challenge to the authors' honesty or to the internal consistency of the synthetic training pipeline; rather, it is a missing domain-transfer check. The proposed negative-control experiment is a direct test: if single-band spectra with realistic self-energies are rarely misclassified, then the concern does not land and the rejection would need to be reconsidered. If they are misclassified, the real-data predictions are ambiguous and the reader's REJECT verdict remains appropriate. No code or data are released, so the test would need to be run by the authors or from a faithful reimplementation.","tokens_in":6114,"tokens_out":4057,"duration_ms":41108,"concrete_test":"Retrain or reuse the CNN with the paper's hyperparameters and evaluate it on a negative-control set of 10,000 synthetic single-band spectra (Mb=0 in Eq. (1)) generated with the KK-consistent self-energy Eq. (7), with P, G, omega0, and noise drawn from the ranges used to model underdoped BSCCO. Measure the fraction classified as 'split' and the regression output d_epsilon_0. If the false-positive rate exceeds 5% or the median predicted d_epsilon_0 exceeds 0.02 eV, the real-data classification is not unambiguously evidence of bilayer splitting; if the false-positive rate is near zero, the reader's rejection should be softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim ('rules out the possibility that bilayer splitting vanishes with underdoping', Conclusions) depends on the CNN outputs for the two spectra in Fig. 8 being causally tied to the presence and magnitude of two dispersions. That link is not tested. The network is trained exclusively on spectra generated by Eqs. (1)-(8), a family in which every image is exactly two Lorentzian bands (or one band after setting Ma or Mb to zero) with the same self-energy, no background, and constant matrix elements. On held-out data from this same generator the classifier reaches only 81% accuracy, so 'unambiguously predicts' already overstates the in-distribution performance. More importantly, no experiment with an independently known splitting value is used to test transfer to real ARPES. Real spectra include a smooth background, photon-energy-dependent and momentum-dependent matrix elements, and stronger self-energy variations than sampled; the underdoped superconducting spectrum in Fig. 8 is precisely the regime where a broad peak-dip-hump from Eq. (7) can mimic or mask a second band. Without a control showing the network does not label single-band, strong-self-energy spectra as split, the 0.10(1)/0.13(1) eV values and the doping trend are unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript applies a convolutional neural network (EfficientNetB0) trained on synthetic ARPES spectra to the question of bilayer splitting in bilayer BSCCO. The synthetic data are generated from a two-band model with a shared self-energy (Eqs. 1-8), whose imaginary part contains a magnetic-resonance peak; the CNN is trained for two tasks: binary classification of the presence of two zones (the 'no-split' class realized by setting one matrix element to zero) and regression of the bare splitting d_epsilon_0. After reporting an in-distribution validation accuracy of 81% for classification and R^2 ~ 0.87 (and 0.80 for a second data set) for regression, the authors apply the network to two real ARPES spectra of optimally doped (x = 0.16, T_c = 90 K) and underdoped (x = 0.10, T_c = 76 K) BSCCO in the superconducting state. The classifier labels both as split and the regression yields d_epsilon_0 = 0.10(1) eV and 0.13(1) eV, from which the authors conclude that bilayer splitting is present across the entire doping range and does not decrease with underdoping, contrary to some theoretical expectations.","tokens_in":6429,"tokens_out":15739,"duration_ms":132608,"significance":"If the claimed result were validated, the finding that bilayer splitting does not shrink on underdoping would settle a long-standing controversy in cuprate photoemission, and the paper would serve as a useful proof-of-concept for training CNN-based spectral analyzers on physically motivated generative models. The authors deserve credit for reporting honest in-distribution metrics (81% classification accuracy; R^2 = 0.87 and 0.80 for regression), for verifying the Kramers-Kronig consistency of the self-energy (Fig. 2 and Eqs. 6-8), and for testing edge cases of weak secondary zones (Fig. 7). The pipeline is falsifiable in principle, since it yields specific quantitative predictions for real spectra that could be checked by independent fitting or by control measurements. However, the physical conclusion as stated is not established by the evidence presented, for the reasons detailed in the major comments below.","major_comments":[{"comment":"The claim that the CNN 'unambiguously predicts the existence of bilayer splitting in all the spectra' and 'rules out the possibility that the bilayer splitting vanishes with underdoping' is not supported by the classifier's reported in-distribution validation accuracy of 81% (Section III): roughly one in five held-out spectra from the same generator is mislabeled, so 'unambiguous' overstates the performance even on the training manifold itself. Because the doping-dependent conclusion rests entirely on the classifier output for the two real spectra in Fig. 8, the authors should at minimum report the classifier's false-positive rate on single-band spectra (one matrix element set to zero) generated with the large resonance strength P and small width G that characterize the underdoped superconducting state, since that is precisely the regime in which a peak-dip-hump line shape can be mistaken for a second band.","section":"§III (Results, Fig. 8); Conclusions"},{"comment":"The transfer of the CNN from the synthetic generator to real ARPES data is asserted but never validated. No spectrum with an independently known splitting value is used to test the network, and the generative family of Eqs. (1)-(8) excludes features of the real spectra it is applied to: no background, no momentum- or photon-energy-dependent matrix elements (Ma and Mb are constant per spectrum), and one shared self-energy for both bands. The 'no-split' class is defined only as exactly one matrix element equal to zero, so the classifier has never seen a realistic single-band spectrum with background and a strong self-energy kink. A concrete, feasible control is to apply the trained classifier to ARPES data of an isostructural single-layer cuprate, or to synthetic spectra augmented with background and out-of-family self-energies; without such a control, the reported 0.10(1) and 0.13(1) eV values remain outputs of the model family rather than measured properties of the real spectra.","section":"§II (Eqs. 1-8); §III (Fig. 8)"},{"comment":"The parenthetical uncertainty in '0.10(1) eV' and '0.13(1) eV' is never defined: the paper describes no ensemble, bootstrap, or cross-validation procedure from which a standard error could be derived. This matters because the text accompanying Fig. 7 states that the regression models 'have some bias near 0.1 eV,' which is exactly the region of the reported optimally doped value, and because the paper's key conclusion — that splitting does not decrease with underdoping — rests entirely on the 0.03 eV difference between the two reported values. Without a defined uncertainty and a quantified bias correction for the regression output, that comparison is not established.","section":"§III (Figs. 7 and 8)"},{"comment":"The conclusion that bilayer splitting is present 'across the entire doping range' and 'does not decrease with underdoping' is drawn from exactly two real samples: one optimally doped (x = 0.16) and one underdoped (x = 0.10), both shown in Fig. 8. Two doping points cannot support a claim about the entire doping range, and the statement that the splitting 'does not decrease' is a claim about a trend that would require at least an additional intermediate doping value; as presented, the evidence shows only that the single underdoped sample produced a larger output value than the single optimally doped sample.","section":"Abstract; Conclusions"},{"comment":"The description of the training distribution is incomplete and internally inconsistent. The sentence 'the training data were restricted to the range of Ma/Mb < 0.1 and Mb/Ma < 0.1' cannot be read literally, since the two inequalities cannot hold simultaneously; the intended allowed range of the matrix-element ratio is therefore unclear, and it is never stated how this range relates to the real spectra in Fig. 8, whose photon energies and hence matrix-element regimes are not given. In addition, the ranges of the self-energy parameters (C, alpha, beta, P, G, omega_0), the hopping t, and the bare splitting d_epsilon_0 used for data generation are not reported. Because the domain of validity of the CNN is exactly its training distribution, and because real spectra include photon-energy regimes in which one band is dominant (cf. Fig. 1, panels f, i, l), the reader cannot assess whether the network operated inside or outside its trained domain when producing the values 0.10 and 0.13 eV.","section":"§III (data generation and training distribution)"}],"minor_comments":[{"comment":"There is a typo in Section I: 'underdoped samo-ple below Tc' should read 'underdoped sample below Tc', and the Section III header 'RESUL TS' is misspelled.","section":"§I; §III heading"},{"comment":"In Eq. (1), the spectral functions are written as A_a(w,k) and A_b(w,k) with the variable w, while Eq. (2) defines A_{a,b}(omega,k); the symbol w appears to be a typo for omega and should be made consistent.","section":"Eq. (1)"},{"comment":"The experimental spectra in Fig. 8 lack provenance details: no reference or dataset identifier for the measurements by the IFW Dresden group at BESSY is given, and the photon energies and measurement geometry are not specified, even though Fig. 1 demonstrates that these strongly affect the visibility of the splitting.","section":"Fig. 8"},{"comment":"The paper provides no code or data availability statement, and the hyperparameter search mentioned in Section II is not reported (the chosen dense-layer sizes, dropout probabilities, and the outcome of the search are not given), which limits reproducibility.","section":"§II"},{"comment":"The phrase 'contrary to some theoretical predictions' does not name the predictions being ruled out; since the paper's main claim is precisely that it contradicts these predictions, the relevant references should be given explicitly.","section":"Conclusions"}],"recommendation":"reject","confidential_remarks":"The central result depends on two real spectra and on a CNN whose transfer to real data is unvalidated; in my view the missing controls (a real single-layer negative control, or at minimum a thorough out-of-family false-positive analysis), the undefined error bars, and the n=2 doping coverage make this a rejection-level problem rather than a minor-revision issue, because the needed fixes require new data or experiments, not just re-analysis. The methodological contribution relative to Refs. [15-17] is incremental, but the physical question is important; if the authors can obtain independent validation and proportionally weaken or re-scope the claims, a resubmission could be competitive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful methodological demonstration that gets oversold as a physics result. The authors train a CNN on spectra generated from a bilayer-split model with a spin-fluctuation self-energy, then apply it to two real BSCCO spectra and conclude that bilayer splitting persists in the underdoped regime. The ML pipeline is sensible, and the Kramers-Kronig consistency check is a nice touch. But the central claim is not supported by the evidence as presented.\n\nWhat's new: using a CNN for this specific classification/regression task on bilayer splitting is new, if incremental over existing ML-for-ARPES work. The hierarchical approach (classify first, then regress) and the explicit comparison of self-energy models without and with the KK-consistent real part are done carefully. The paper is honest that training data are modeled, and it flags the low-intensity limitation.\n\nWhere it falls short: the network is never validated on real spectra with independently known splitting. The only test is held-out synthetic data from the same generator, and the classifier gets 81% accuracy there, which undercuts 'unambiguously predicts.' The regression errors in Fig. 6 are substantial near the cutoff. When applied to two real spectra (one optimal, one underdoped), the outputs are reported as 0.10(1) and 0.13(1) eV with no statement about how those uncertainties were derived. The claim about 'the entire doping range' is drawn from two samples. And the key worry is the one the stress-test note identifies: the generative model family always contains exactly two Lorentzian bands (or one), with the same functional self-energy and no background. A network trained on that family may be reacting to peak shape or broad spectral features rather than two well-defined bands. Without a negative control (e.g., synthetic single-band spectra with the same strong self-energy, or real spectra where splitting is absent), the doping trend is not established.\n\nThe citation pattern is fine; the earlier references by the same group did claim the same result, so the paper is a confirmation attempt, not a new discovery. That is okay if the method adds value, but it also means the physics conclusion is not novel.\n\nWho this is for: someone working on ML-assisted ARPES analysis might get ideas from the pipeline, and a referee should ask for the missing validation. As presented, I would not cite the physics conclusion. If the authors add a proper generalization test, release code and data, and temper the abstract, this could become a solid methods paper.\n\nRecommendation: send to peer review but prepare for heavy revision; desk rejection would be too harsh given the careful synthetic modeling and the clear statement of the approach, but the central claim as written should not survive unchanged.","headline":"A careful synthetic-data CNN study whose central physics claim is unsupported by validation on real spectra; the pipeline is worth a look, the conclusion is not.","tokens_in":6926,"tokens_out":2496,"would_cite":false,"duration_ms":22885,"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":"A convolutional neural network trained purely on synthetic photoemission spectra detects bilayer splitting in real BSCCO data and finds it persists at 0.13(1) eV in underdoped samples, ruling out its disappearance with underdoping.","keywords":["bilayer splitting","cuprate superconductors","ARPES","convolutional neural network","self-energy","Kramers-Kronig relations","machine learning","BSCCO"],"falsifier":"Measure the same two samples' band splitting with an independent method that does not rely on the paper's assumptions—for example, by fitting the raw spectrum's peak positions directly or using a different theoretical model—and check whether the values agree with $0.10(1)$ eV and $0.13(1)$ eV. Another check: apply the network to one sample measured at several photon energies; the predicted splitting should be the same (within $0.01$ eV) at every energy, because real splitting does not depend on photon energy.","tokens_in":5900,"feed_emoji":"🤖","tokens_out":13917,"duration_ms":104772,"temperature":0.7,"pith_summary":"This paper tries to settle a long-standing controversy over whether the bilayer splitting—the energy separation between the two bands formed by the two CuO layers in a bilayer cuprate—persists in the underdoped superconducting state. The authors argue that a convolutional neural network trained only on synthetic photoemission spectra can separate the coherent bilayer splitting from the incoherent broadening produced by magnetic fluctuations, and can read the splitting energy out of real experimental spectra. Applied to BSCCO, the network returns $0.10(1)$ eV for optimally doped samples and $0.13(1)$ eV for underdoped samples, leading to the conclusion that the splitting does not vanish or shrink with underdoping. The broader claim is that machine learning trained on physically motivated models is a workable tool for disentangling overlapping contributions in correlated-electron spectroscopy.","feed_headline":"Bilayer splitting survives underdoping, CNN analysis finds","feed_subtitle":"A CNN trained on synthetic spectra reads 0.10 eV and 0.13 eV splits in optimally and underdoped BSCCO.","key_machinery":"The central machinery is a generative model of the photoemission intensity as the sum of two spectral functions for the antibonding and bonding bands (the two bands formed by the coupling of the two CuO layers in a bilayer), each with a simple cosine dispersion and sharing a single complex self-energy $\\Sigma(\\omega)$. The imaginary part of $\\Sigma$ contains a Lorentzian resonance peak simulating the magnetic mode, and the real part is either omitted or obtained analytically from Kramers-Kronig relations. A convolutional neural network with an EfficientNet backbone is trained on tens of thousands of synthetic images of these spectra, with added noise and matrix-element variations, first to classify the presence of splitting and then to regress the splitting energy $d\\varepsilon_0 = \\varepsilon_0^a - \\varepsilon_0^b$. The trained network is applied to experimental spectra through a hierarchical dual-model approach: the classification model gates whether splitting is present, and the regression model outputs the splitting value.","core_discovery":"The central claim is that the bilayer splitting in BSCCO exists across the entire doping range, including severely underdoped samples in the superconducting state, and that its magnitude does not decrease with underdoping. The authors train a CNN on thousands of spectra generated from a two-band model in which the two split bands share a common complex self-energy $\\Sigma(\\omega)$—the energy-dependent correction that broadens and shifts the spectral lines—including a resonance peak that mimics magnetic fluctuations; the model is built with and without a Kramers-Kronig-consistent real part. After validation on synthetic data, the network is applied to real ARPES spectra from optimally doped ($T_c=90$ K, $x=0.16$) and underdoped ($T_c=76$ K, $x=0.10$) BSCCO. The classification stage unambiguously predicts bilayer splitting in all experimental spectra, and the regression stage gives splitting values $0.10(1)$ eV and $0.13(1)$ eV, respectively. From this the authors conclude that the bilayer splitting does not vanish with underdoping.","pith_inferences":["The absolute splitting values depend on how faithfully the generative model captures real self-energy and matrix-element effects; a natural check is whether the predicted $d\\varepsilon_0$ for one sample is stable across different photon energies, since the physical splitting should be photon-energy independent while the matrix elements in the model are not.","The qualitative trend—no decrease of splitting with underdoping—is probably more robust than the absolute numbers, because any systematic bias from the self-energy parameterization would affect both doping levels similarly; this makes the qualitative conclusion the paper's stronger result.","The framework could be extended to map the full doping-temperature phase diagram of bilayer cuprates with a single trained network, or adapted to separate bilayer splitting from pseudogap effects and other self-energy structures in the same spectra."],"forward_implications":["The bilayer splitting is present in BSCCO at optimal doping and in the severely underdoped superconducting state, with inferred values $0.10(1)$ eV and $0.13(1)$ eV.","The magnitude of the splitting does not decrease with underdoping, so theories of cuprates that predict a vanishing or strongly suppressed interlayer coupling in the underdoped regime would need to be revised.","Training CNNs on physically modeled spectra can substitute for large experimental training sets, offering a general route for extracting parameters from ARPES and similar spectroscopies.","The classification-then-regression pipeline can be reused for other spectral-feature extraction problems where coherent and incoherent contributions overlap."],"supporting_citations":[{"why":"First observation of bilayer splitting in overdoped BSCCO, establishing the phenomenon whose doping dependence this paper contests.","marker":"[7]"},{"why":"ARPES spectra of bilayer BSCCO at various photon energies that illustrate where the splitting is hard to see in underdoped superconducting samples.","marker":"[10]"},{"why":"Prior evidence of bilayer splitting and coherence effects in optimal and underdoped BSCCO that this paper confirms and extends.","marker":"[11]"},{"why":"Supplies the spin-fluctuation resonance form of the self-energy used to generate the synthetic spectra.","marker":"[12]"},{"why":"Provides experimental constraints on the spin-fluctuation coupling strength used in the self-energy parameterization.","marker":"[13]"},{"why":"Establishes the self-consistent self-energy analysis and the high-energy-tail correction used in the Kramers-Kronig-consistent model.","marker":"[14]"},{"why":"The EfficientNet architecture that serves as the backbone of the convolutional neural network.","marker":"[19]"}],"fun_headline_variants":["CNN confirms bilayer splitting persists across full doping range","Bilayer splitting survives underdoping, machine learning finds","Stable bilayer splitting seen in BSCCO by CNN at all dopings","Underdoping doesn't kill bilayer splitting, CNN analysis shows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic spectra used to train the network faithfully mimic real photoemission measurements—including the broadening from magnetic fluctuations and ignoring background signals and momentum-dependent measurement factors—so that a network trained only on synthetic data can correctly read the band splitting in real data.","fun_headline_variants_meta":{"raw":{"variants":["CNN confirms bilayer splitting persists across full doping range","Bilayer splitting survives underdoping, machine learning finds","Stable bilayer splitting seen in BSCCO by CNN at all dopings","Underdoping doesn't kill bilayer splitting, CNN analysis shows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1647,"prompt_tokens":907,"completion_tokens":740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":669}},"tokens_in":523,"tokens_out":740,"duration_ms":6609,"temperature":1.0,"reasoning_tokens":669,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:15:59.009452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same two samples' band splitting with an independent method that does not rely on the paper's assumptions—for example, by fitting the raw spectrum's peak positions directly or using a different theoretical model—and check whether the values agree with $0.10(1)$ eV and $0.13(1)$ eV. Another check: apply the network to one sample measured at several photon energies; the predicted splitting should be the same (within $0.01$ eV) at every energy, because real splitting does not depend on photon energy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"First observation of bilayer splitting in overdoped BSCCO, establishing the phenomenon whose doping dependence this paper contests."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ARPES spectra of bilayer BSCCO at various photon energies that illustrate where the splitting is hard to see in underdoped superconducting samples."},{"cited_title":"Chuang, A","cited_arxiv_id":null,"evidence_quote":"Prior evidence of bilayer splitting and coherence effects in optimal and underdoped BSCCO that this paper confirms and extends."},{"cited_title":"Eschrig, The effect of collective spin-1 excitations on electronic spectra in high- tc superconductors, Advances in Physics 55, 47 (2006)","cited_arxiv_id":null,"evidence_quote":"Supplies the spin-fluctuation resonance form of the self-energy used to generate the synthetic spectra."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides experimental constraints on the spin-fluctuation coupling strength used in the self-energy parameterization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the self-consistent self-energy analysis and the high-energy-tail correction used in the Kramers-Kronig-consistent model."}],"review_version":1}