Disentangling Coherent and Incoherent Effects in Superconductor Photoemission Spectra via Machine Learning

A. A. Kordyuk; K. H. Bohachov

REVIEW 5 major objections 5 minor 21 references

Disentangling Coherent and Incoherent Effects in Superconductor Photoemission Spectra via Machine Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2412.11129 v1 pith:KEZWLEZA submitted 2024-12-15 cond-mat.supr-con cond-mat.str-el

K. H. Bohachov , A. A. Kordyuk This is my paper · ORCID

classification cond-mat.supr-concond-mat.str-el

keywords bilayersplittingcupratesuperconductorsARPESconvolutionalneuralnetworkself-energyKramers-KronigrelationsmachinelearningBSCCO

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

What'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.

Where 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.

The 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.

Who 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.

Recommendation: 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.

Referee Report

5 major / 5 minor

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.

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 (5)

[§III (Results, Fig. 8); Conclusions] 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.
[§II (Eqs. 1-8); §III (Fig. 8)] 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.
[§III (Figs. 7 and 8)] 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.
[Abstract; Conclusions] 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.
[§III (data generation and training distribution)] 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.

minor comments (5)

[§I; §III heading] 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.
[Eq. (1)] 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.
[Fig. 8] 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.
[§II] 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.
[Conclusions] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNN prediction is model-dependent but not equivalent to its inputs by construction.

full rationale

The paper's derivation chain is: (i) generate synthetic ARPES spectra from a two-band model with a resonance self-energy, labeling each image by the input splitting dε0 or by one-band/two-band class; (ii) train a CNN to classify/regress that label; (iii) apply the trained CNN to two real BSCCO spectra and read off predicted splittings; (iv) conclude that bilayer splitting persists and does not decrease with underdoping. There is no step in which the predicted quantity is defined from the real data's own labels, and no parameter is fitted to the real spectra and then renamed as a prediction. The CNN weights are learned from simulated images, not from the experimental spectra, so the 0.10(1) eV and 0.13(1) eV outputs are not forced by construction. The fact that dε0 is defined by Eqs. (1)-(3) is simply the choice of supervised target; the generative model also includes one-band examples (Ma or Mb set to zero), so the classifier is not tautologically positive. The main weakness is external validity: the network is never tested on real spectra with independently known splitting, and the 81% validation accuracy shows that even in-distribution performance is imperfect. Those are correctness and generalizability concerns, not circularity. The self-citations (refs. [9]-[11], [14]) are prior experimental/analysis work used as context or empirical input, not as the sole justification of the conclusion; the conclusion rests on the CNN output for real data. Therefore no circularity score above 0 is warranted.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends entirely on the generative model in Eqs. (1-8). The CNN is trained on this model and then applied to real data; any conclusion is therefore conditional on the model family being a faithful representation of the experimental spectra. The paper does not validate this assumption against spectra with known splitting.

free parameters (6)

Self-energy parameters (C, alpha, beta, P, G, omega0) = not disclosed
Used in Eq. (4)/(7) to generate training spectra; ranges not given in the paper, chosen to match earlier experiments by the same group.
Hopping t in dispersions = not disclosed
Defines band dispersions in Eq. (3); training range not given.
Bare splitting d_epsilon0 range = not disclosed
Target label for regression; examples 8, 21, 89 meV shown, full range not stated.
Matrix element ratios Ma/Mb = restricted to <0.1 or >10
Training restricted to ratios where one band dominates; this selection improves R2 by 0.08 but biases the model.
Gaussian noise standard deviation = [0,50] in arbitrary units
Chosen by analyzing numerous spectra; intensity scale not defined.
CNN hyperparameters = not fully specified
Dense layers, dropout, backbone found by hyperparameter search; only final architecture described.

assumptions (4)

domain assumption ARPES intensity is a sum of two independent split bands with a common self-energy
Eq. (1) assumes no background, no interference between bands, and no matrix element momentum dependence beyond Ma/Mb.
ad hoc to paper The self-energy form of Eq. (4)/(7) captures all relevant incoherent effects in the cuprates
The magnetic resonance peak is modeled as a Lorentzian; other incoherent processes are absorbed into C, alpha, beta terms.
ad hoc to paper A CNN trained on synthetic spectra from this model generalizes to real experimental spectra
No out-of-distribution validation is performed; real data may contain backgrounds, experimental artifacts, and self-energy shapes outside the training family.
domain assumption Fermi-Dirac distribution and a single k-axis dispersion with cosine form suffice
Eq. (3) uses a simple cosine band, ignoring bilayer momentum dependence of matrix elements and possible bilayer-dependent self-energies.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Disentangling Coherent and Incoherent Effects in Superconductor Photoemission Spectra via Machine Learning." pith.science (2026). https://pith.science/paper/KEZWLEZA

@misc{pith2026241211129,
  author       = {Pith},
  title        = {Pith review of: Disentangling Coherent and Incoherent Effects in Superconductor Photoemission Spectra via Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KEZWLEZA}},
  note         = {Machine review of arXiv:2412.11129}
}

read the original abstract

Disentangling coherent and incoherent effects in the photoemission spectra of strongly correlated materials is generally a challenging problem due to the involvement of numerous parameters. In this study, we employ machine learning techniques, specifically Convolutional Neural Networks (CNNs), to address the long-standing issue of the bilayer splitting in superconducting cuprates. We demonstrate the effectiveness of CNN training on modeled spectra and confirm earlier findings that establish the presence of bilayer splitting across the entire doping range. Furthermore, we show that the magnitude of the splitting does not decrease with underdoping, contrary to expectations. This approach not only highlights the potential of machine learning in tackling complex physical problems but also provides a robust framework for advancing the analysis of electronic properties in correlated superconductors.

Figures

Figures reproduced from arXiv: 2412.11129 by the authors.

**Figure 2.** FIG. 2. Modeled self-energy parts (bottom): imaginary (red), [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** FIG. 3. Model architecture [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 6.** Figure 6: FIG. 6. The distribution of absolute errors in percentage for [PITH_FULL_IMAGE:figures/full_fig_p003_6.png]

**Figure 7.** Figure 7: FIG. 7. a) Absolute error of [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

**Figure 8.** Figure 8: FIG. 8. Real data bilayer splitting prediction using a hier [PITH_FULL_IMAGE:figures/full_fig_p004_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 15 canonical work pages

[1]

P. W. Anderson, The Resonating Valence Bond State in La 2CuO4 and Superconductivity, Science 235, 1196 (1987)

1987
[2]

R. Liu, B. W. Veal, A. P. Paulikas, J. W. Downey, P. J. Kosti´ c, S. Fleshler, U. Welp, C. G. Olson, X. Wu, A. J. Arko, and J. J. Joyce, Fermi-surface topology of yba2cu3ox with varied oxygen stoichiometry: A photoe- mission study, Phys. Rev. B 46, 11056 (1992)

work page 1992
[3]

Andersen, A

O. Andersen, A. Liechtenstein, O. Jepsen, and F. Paulsen, LDA energy bands, low-energy Hamiltonians, t′, t′′, t⊥(k), and J⊥J, Journal of Physics and Chemistry of Solids 56, 1573 (1995)

work page 1995
[4]

J. M. Luttinger and J. C. Ward, Ground-state energy of a many-fermion system. ii, Phys. Rev. 118, 1417 (1960)

1960
[5]

H. Ding, A. F. Bellman, J. C. Campuzano, M. Randeria, M. R. Norman, T. Yokoya, T. Takahashi, H. Katayama- Yoshida, T. Mochiku, K. Kadowaki, G. Jennings, and G. P. Brivio, Electronic Excitations in Bi 2Sr2CaCu2O8: Fermi Surface, Dispersion, and Absence of Bilayer Split- ting, Phys. Rev. Lett. 76, 1533 (1996)

work page 1996
[6]

P. W. Anderson, The Theory of Superconductivity in the High-Tc Cuprate Superconductors(1997)

work page 1997
[7]

D. L. Feng, N. P. Armitage, D. H. Lu, A. Damascelli, J. P. Hu, P. Bogdanov, A. Lanzara, F. Ronning, K. M. Shen, H. Eisaki, C. Kim, Z.-X. Shen, J.-i. Shimoyama, and K. Kishio, Bilayer Splitting in the Electronic Structure of Heavily Overdoped Bi 2Sr2CaCu2O8+δ, Phys. Rev. Lett. 86, 5550 (2001)

work page 2001
[8]

Chuang, A

Y.-D. Chuang, A. D. Gromko, A. Fedorov, Y. Aiura, K. Oka, Y. Ando, H. Eisaki, S. I. Uchida, and D. S. Dessau, Doubling of the Bands in Overdoped Bi2Sr2CaCu 2O8+δ: Evidence for c-Axis Bilayer Cou- pling, Phys. Rev. Lett. 87, 117002 (2001)

work page 2001

Show all 21 references

[9]

A. A. Kordyuk, S. V. Borisenko, M. S. Golden, S. Legner, K. A. Nenkov, M. Knupfer, J. Fink, H. Berger, L. Forr´ o, and R. Follath, Doping dependence of the Fermi surface in Bi2Sr2CaCu2O8+δ, Phys. Rev. B 66, 014502 (2002)

2002
[10]

S. V. Borisenko, A. A. Kordyuk, T. K. Kim, A. Koitzsch, M. Knupfer, J. Fink, M. S. Golden, M. Eschrig, H. Berger, and R. Follath, Anomalous enhancement of 6 the coupling to the magnetic resonance mode in under- doped Pb-Bi2212, Phys. Rev. Lett. 90, 207001 (2003)

2003
[11]

Chuang, A

Y.-D. Chuang, A. D. Gromko, A. V. Fedorov, Y. Aiura, K. Oka, Y. Ando, M. Lindroos, R. S. Markiewicz, A. Ban- sil, and D. S. Dessau, Bilayer Splitting and Coherence Effects in Optimal and Underdoped Bi2Sr2CaCu 2O8+δ, Phys. Rev. B 69, 094515 (2004)

2004
[12]

Eschrig, The effect of collective spin-1 excitations on electronic spectra in high- tc superconductors, Advances in Physics 55, 47 (2006)

M. Eschrig, The effect of collective spin-1 excitations on electronic spectra in high- tc superconductors, Advances in Physics 55, 47 (2006)

2006
[13]

T. Dahm, V. Hinkov, S. V. Borisenko, A. A. Ko- rdyuk, V. B. Zabolotnyy, J. Fink, B. Buchner, D. J. Scalapino, W. Hanke, and B. Keimer, Strength of the spin-fluctuation-mediated pairing interaction in a high- temperature superconductor, Nat. Phys. 5, 217 (2009)

2009
[14]

A. A. Kordyuk, S. V. Borisenko, A. Koitzsch, J. Fink, M. Knupfer, and H. Berger, Bare electron dispersion from experiment: Self-consistent self-energy analysis of pho- toemission data, Phys. Rev. B 71, 214513 (2005)

2005
[15]

S. A. Ekahana, G. I. Winata, Y. Soh, A. Tamai, R. Mi- lan, G. Aeppli, and M. Shi, Transfer learning application of self-supervised learning in arpes, Machine Learning: Science and Technology 4, 035021 (2023)

2023
[16]

T. K. Imamura, M., Unsupervised learning of spatially- resolved arpes spectra for epitaxially grown graphene via non-negative matrix factorization, Sci Rep 14, 24200 (2024)

2024
[17]

J. Liu, D. Huang, Y.-f. Yang, and T. Qian, Removing grid structure in angle-resolved photoemission spectra via deep learning method, Phys. Rev. B 107, 165106 (2023)

2023
[18]

K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learning for image recognition, CoRR abs/1512.03385 (2015), 1512.03385

2015 arXiv
[19]

Tan and Q

M. Tan and Q. V. Le, Efficientnet: Rethinking model scaling for convolutional neural networks, CoRR abs/1905.11946 (2019), 1905.11946

2019 arXiv
[20]

G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, Improving neural networks by preventing co-adaptation of feature detectors, CoRR abs/1207.0580 (2012), 1207.0580

2012 arXiv
[21]

D. P. Kingma and J. Ba, Adam: A method for stochastic optimization (2017), arXiv:1412.6980 [cs.LG]

2017 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

P. W. Anderson, The Resonating Valence Bond State in La 2CuO4 and Superconductivity, Science 235, 1196 (1987)

1987

[2] [2]

R. Liu, B. W. Veal, A. P. Paulikas, J. W. Downey, P. J. Kosti´ c, S. Fleshler, U. Welp, C. G. Olson, X. Wu, A. J. Arko, and J. J. Joyce, Fermi-surface topology of yba2cu3ox with varied oxygen stoichiometry: A photoe- mission study, Phys. Rev. B 46, 11056 (1992)

work page 1992

[3] [3]

Andersen, A

O. Andersen, A. Liechtenstein, O. Jepsen, and F. Paulsen, LDA energy bands, low-energy Hamiltonians, t′, t′′, t⊥(k), and J⊥J, Journal of Physics and Chemistry of Solids 56, 1573 (1995)

work page 1995

[4] [4]

J. M. Luttinger and J. C. Ward, Ground-state energy of a many-fermion system. ii, Phys. Rev. 118, 1417 (1960)

1960

[5] [5]

H. Ding, A. F. Bellman, J. C. Campuzano, M. Randeria, M. R. Norman, T. Yokoya, T. Takahashi, H. Katayama- Yoshida, T. Mochiku, K. Kadowaki, G. Jennings, and G. P. Brivio, Electronic Excitations in Bi 2Sr2CaCu2O8: Fermi Surface, Dispersion, and Absence of Bilayer Split- ting, Phys. Rev. Lett. 76, 1533 (1996)

work page 1996

[6] [6]

P. W. Anderson, The Theory of Superconductivity in the High-Tc Cuprate Superconductors(1997)

work page 1997

[7] [7]

D. L. Feng, N. P. Armitage, D. H. Lu, A. Damascelli, J. P. Hu, P. Bogdanov, A. Lanzara, F. Ronning, K. M. Shen, H. Eisaki, C. Kim, Z.-X. Shen, J.-i. Shimoyama, and K. Kishio, Bilayer Splitting in the Electronic Structure of Heavily Overdoped Bi 2Sr2CaCu2O8+δ, Phys. Rev. Lett. 86, 5550 (2001)

work page 2001

[8] [8]

Chuang, A

Y.-D. Chuang, A. D. Gromko, A. Fedorov, Y. Aiura, K. Oka, Y. Ando, H. Eisaki, S. I. Uchida, and D. S. Dessau, Doubling of the Bands in Overdoped Bi2Sr2CaCu 2O8+δ: Evidence for c-Axis Bilayer Cou- pling, Phys. Rev. Lett. 87, 117002 (2001)

work page 2001