Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Transfer learning electronic structure: millielectron volt accuracy for sub-million-atom moir\'e semiconductor

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Transfer learning predicts sub-meV moiré Hamiltonians at 253k atoms

desk verdict Solid DFT-validated core down to 1.89°, but the headline sub-meV claim below that rests on an unverified extrapolation; still deserves serious refereeing. read the letter →

arxiv 2501.12452 v1 pith:VDQDUVD6 submitted 2025-01-21 cond-mat.mtrl-sci cond-mat.str-el

Ting Bao , Ning Mao , Wenhui Duan , Yong Xu , Adrian Del Maestro , Yang Zhang This is my paper · ORCID
classification cond-mat.mtrl-scicond-mat.str-el
keywords transferlearningneuralnetworkelectronicstructuremoirématerialstwistedMoTe2Chernnumberdensityfunctionaltheorylinearscaling
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

The paper claims that a neural network for electronic Hamiltonians can be made accurate for twisted (moiré) semiconductors without training on the expensive small-angle structures that are hardest to compute. By first pre-training on ordinary non-twisted MoTe2 and then fine-tuning on just 48 large-angle twisted configurations, the network predicts Hamiltonian matrix elements with mean absolute errors below 0.1 meV for a 1000-atom system, and it scales linearly to a 253,140-atom nanoribbon. On that ultra-large system it reproduces edge states whose Chern numbers match the predicted bulk topology. If the claim holds, transfer learning turns machine-learned electronic structure into a practical tool for moiré materials at device-relevant sizes.

What carries the argument

The load-bearing object is the two-step transfer learning scheme applied to an equivariant message-passing neural network that outputs sparse Hamiltonian matrix elements in a pseudo-atomic-orbital basis, including spin-orbit coupling. Pre-training on non-twisted structures teaches generic hopping and overlap terms; fine-tuning on 48 large-angle twisted structures teaches the long-range, Hartree-like contributions that dominate small-angle moiré physics. The paper then extracts band topology from the predicted Hamiltonian using C3 symmetry eigenvalues at high-symmetry points, and uses partial sparse diagonalization to achieve O(N) scaling for edge-state projections.

What would settle it

Run a DFT calculation for twisted MoTe2 at a small angle not included in the training set, such as 1.2° or 1.5°, compare the first few band energies with the neural network prediction; if the mean absolute error of the top three bands exceeds roughly 1 meV, the claimed sub-meV accuracy at small angles would be falsified. A second test is to compute the same comparison for a strained structure at a twist angle below 2°, where the model has no direct training data.

Watch

Extended reading notes

Core claim

The central claim is that a two-step transfer learning procedure—pre-training on 576 non-twisted MoTe2 supercells and fine-tuning on 48 twisted structures with twist angles between 9° and 21°—yields an equivariant neural network that predicts the full spin-orbit-coupled Hamiltonian of twisted MoTe2 in a localized-orbital basis with sub-meV elementwise accuracy across twist angles from 5.08° down to 1.89°, validated against DFT. The same model extrapolates to smaller angles (down to 0.88°) and to strained and nanoribbon geometries, where it predicts Chern numbers from C3 eigenvalues and edge states consistent with those invariants. The authors frame this as a resolution of the accuracy-efficiency tradeoff: twisted-only training is too costly, non-twisted-only training is too inaccurate for long-wavelength moiré effects, and the transfer step captures the missing long-range physics with minimal extra data.

Load-bearing premise

The model is fine-tuned on only 48 twisted structures with twist angles between 9° and 21°, yet it is used to predict Hamiltonians at 1.89° and below, down to 0.88°, where no DFT reference is provided, so the load-bearing premise is that long-range interactions learned from large-angle twisted structures extrapolate smoothly to the much longer moiré wavelengths of small-angle structures.

Editorial extensions

If this is right

  • Electronic structure calculations for moiré systems with millions of atoms become feasible on a single workflow, since both inference and diagonalization scale linearly with atom count.
  • The predicted Hamiltonians can serve as input to many-body or transport calculations, enabling correlated-phase studies at realistic device scales.
  • Strained twisted systems, which break the C3 symmetry used in training, are still captured accurately enough to reproduce strain-induced topological transitions.
  • The same transfer learning recipe should apply to other twisted van der Waals materials, potentially reducing the cost of exploring their phase diagrams.

Reading between the lines

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

  • The extrapolation from 9–21° fine-tuning data to sub-2° structures is the least directly tested link; a DFT benchmark at 1.0–1.5° would either confirm the learned long-range extrapolation or reveal an error floor.
  • Because the model relies on C3 eigenvalues for topology, it inherits the assumption that the relevant bands remain isolated at the high-symmetry points; at the smallest angles, where bands become nearly flat and mixed, Chern number assignment may become ambiguous.
  • The claimed O(N) scaling depends on the sparsity of the Hamiltonian and the choice of a narrow energy window for diagonalization; for very dense band crossings the required eigenvalue count could grow, degrading the practical scaling.
  • A natural extension is to use the same transfer learning to predict not just the Hamiltonian but also response functions or Berry curvature, which would test whether sub-meV Hamiltonian accuracy translates into quantitatively correct topological markers.
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.

Referee Report

3 major / 4 minor

Summary. The paper presents a transfer learning scheme for predicting tight-binding-like PAO Hamiltonians of twisted bilayer MoTe2, combining DeepH-E3 equivariant graph neural networks with pre-training on non-twisted structures and fine-tuning on 48 large-angle twisted structures. The authors report mean absolute errors below 0.1 meV in Hamiltonian matrix elements for twist angles from 5.08° to 1.89° when validated against DFT, correct topological Chern numbers derived from C3 eigenvalues over the same range, and application to strained moiré systems. They then extrapolate to smaller twist angles down to 0.88°, where no DFT reference is provided, and use the model to compute band structures and edge states of twisted nanoribbons with up to 253,140 atoms (8,775,520 orbitals), claiming O(N) scaling of inference and partial diagonalization. The central accuracy claims are grounded in direct DFT comparisons for the 5.08°-1.89° range, while the sub-million-atom nanoribbon and all angles below 1.89° rest on a smooth and untested extrapolation of long-range (Hartree-like) interactions across more than an order-of-magnitude change in moiré wavelength.

Significance. If the claimed sub-meV Hamiltonian accuracy and O(N) scalability hold at small twist angles, this would be a practically important step for simulating realistic large-scale moiré devices. The paper's strengths include a direct elementwise comparison of Hamiltonian matrix blocks to OpenMX DFT results, a clear demonstration that transfer learning improves accuracy compared to a model trained only on non-twisted structures, validated Chern numbers across ten twist angles, and a promising application to strained systems. The reproducibility details in the Supplemental Material (training set sizes, hyperparameters, basis sets, computational tables) are unusually complete for a machine-learning electronic-structure paper. The main weakness is that the headline 'millielectron volt accuracy' for sub-million-atom and small-angle systems is not supported by any DFT benchmark below 1.89°, nor for the nanoribbon geometry, so the significance of the flagship result is currently conditional on an untested extrapolation premise.

major comments (3)
  1. [Sec. VI, Fig. 4(c), Supp. Figs. S4/S5, Supp. Table S7] All predictions below 1.89° (1.70°, 1.47°, 1.30°, 1.08°, 0.88°) lack any DFT reference, and the nanoribbon edge-state claim in Fig. 4(c) is only checked for consistency with Chern numbers computed from the same neural-network Hamiltonian. Since the abstract and conclusion claim 'millielectron volt accuracy' and accurate sub-million-atom edge states, this extrapolation is load-bearing. Please provide at least one DFT benchmark at an intermediate small angle (e.g., 1.70° or 1.47°) for bulk tMoTe2, and an independent validation of the nanoribbon edge states (for example, a DFT or continuum-model comparison for a smaller ribbon, or a converged large-angle ribbon with known edge physics). Without such checks, the small-angle and nanoribbon claims should be presented as extrapolations rather than validated predictions.
  2. [Supp. Tables S1/S2, Fig. S3(c), Fig. 4(b)] The twist angle 2.00° is a clear outlier within the DFT-validated range: in Table S1 the Mo-Mo average error after transfer learning is 0.14 meV versus 0.07 meV for neighboring angles, and in Table S2 the Mo-Mo maximum error is 2.73 meV versus roughly 1.2-1.6 meV at adjacent angles, with much smaller percentage reductions after transfer learning. The paper does not explain this anomaly, yet the 2.00° data point is included in the MAE trend of Fig. 4(b). Please clarify whether this indicates a genuine difficulty at a particular commensurate angle, a data or training issue, or a model limitation; this is important for judging the reliability of the angle dependence the paper reports.
  3. [Supp. C, Secs. III-IV] The fine-tuning dataset contains only 48 twisted structures with twist angles between 9° and 21°, so even the validated range 5.08°-1.89° is an extrapolation of more than a factor of four in moiré wavelength, and the range below 1.89° extends this by another factor of two. The paper itself identifies the pre-transfer-learning error growth at small angles as 'arising from a difficulty to describe long-range effects like the Hartree term' (Sec. III), yet it provides no direct test of whether the transfer-learned model extrapolates this long-range contribution correctly to the 0.88° case (moiré wavelength about 200 Å). Please add a quantitative assessment of angle extrapolation, for example by training on a subset of the available DFT-validated angles (e.g., only angles > 3.5°) and reporting how the MAE degrades at 2.13° and 1.89°; this would provide direct evidence for or against the smooth-transfer premise that currently underpins the small-angle and nanoribbon claims.
minor comments (4)
  1. [Sec. II] Equation (1) writes Hiα,jβ(k)ψnk = En(k)Siα,jβ(k)ψnk; the band index on the eigenvector and eigenvalue is inconsistent (ψnk appears on both sides without a clear summation over jβ). Please make the generalized eigenvalue equation notation explicit.
  2. [Sec. V] The text repeatedly says 'stained' where 'strained' is meant (e.g., 'predictions for stained tMoTe2'). Please correct the typo throughout.
  3. [Sec. VI and Supp. Table S6] The memory figure '2142 GB' for the 10× nanoribbon (253,140 atoms) appears extremely large for the given matrix size (8,775,520 orbitals) and may be in units of MB or include a factor that is not defined. Please verify the units and clarify how the memory scales with system size in relation to the claimed O(N) behavior.
  4. [Fig. 4 caption] The caption says the nanoribbon is periodic along lattice vector b and finite along a, spanning a length of 5 unit cells, but the text in Sec. VI says '10 times of a single 0.88°-twisted unit cell' for the largest system. Please clarify which quantity is a multiple of the unit cell and which is the ribbon length.

Circularity Check

1 steps flagged · score 3.0 of 10

The DFT-benchmarked Hamiltonian accuracy claim is independent and non-circular; the 0.88° nanoribbon edge-state check is an internal self-consistency loop.

  1. other [Section VI (Scalability Performance), nanoribbon edge-state paragraph; Fig. 4(c); Table S7 (0.89°)]
    "The edge states align with the Chern number of the periodic 0.88◦-twisted structures. Specifically, gapless edge states are observed between the original first band (Chern number -1) and the second band (Chern number 1), whereas gapped edge states are present between the original second band (Chern number 1) and the third band (Chern number 0)."

    The 0.89° Chern numbers in Table S7 and the Fig. 4(c) edge-state projection are both computed from the same transfer-learning Hamiltonian; no DFT reference exists below 1.89° (Figs. S4–S5 are NN predictions only). Bulk-boundary correspondence makes edge-state/bulk-Chern agreement a theorem for any Hamiltonian, so the observed consistency is forced by construction and cannot certify the accuracy of the NN Hamiltonian. The phrase 'accurately capturing edge states' is therefore a self-consistency loop, not an independent validation.

full rationale

The central accuracy claim is not circular: the fine-tuned network is benchmarked against DFT (OpenMX) Hamiltonian matrix elements and band structures for angles from 5.08° to 1.89° (Tables S1–S2, Figs. S1–S3, S6–S11), and these angles lie below the 9°–21° fine-tuning set, so the sub-meV errors are genuine out-of-sample predictions, not fitted values. The Chern-number sequence over 5.08°–2.13° is also checked against DFT and consistent with experiments. The one circular element is the 0.88° nanoribbon edge-state demonstration, where both the edge states and the 'predicted Chern numbers' used as reference are outputs of the same NN Hamiltonian, making the agreement an internal consistency check rather than external validation. Small-angle predictions below 1.89° are extrapolations with no DFT reference; this is a correctness risk (the paper itself notes error growing with decreasing angle 'arising from a difficulty to describe long-range effects like the Hartree term'), not circularity. Self-citations (DeepH-E3, DDHT) supply architecture and pretraining details but are not load-bearing for the accuracy benchmark. Score reflects one peripheral self-consistency loop while the main derivation is DFT-anchored.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard DFT reference data, a chosen basis and cutoff, and the key transfer learning assumption that large-angle twisted structures teach the model long-range physics applicable to much smaller angles. No new physical entities are introduced.

free parameters (2)
  • Hamiltonian hopping cutoff radius Rc
    The sparse Hamiltonian is built from hopping elements within a cutoff radius Rc; the value is not stated in the paper, and the central accuracy claim depends on this truncation. This is a modeling choice chosen by hand.
  • PAO basis cutoff radius (7.0 Bohr) = 7.0 Bohr
    Basis sets Mo7.0-s3p2d1 and Te7.0-s3p2d2 fix the orbital cutoff at 7.0 Bohr, chosen by hand; the Hamiltonian quality depends on this choice.
assumptions (4)
  • domain assumption PBE functional and norm-conserving pseudopotentials yield accurate reference Hamiltonians and overlap matrices for twisted MoTe2.
    The entire training target comes from OpenMX PBE calculations; if PBE is inaccurate for these small-bandwidth systems, the model inherits the error. Invoked in Sec. II and Supplemental B.
  • domain assumption The DeepH-E3 architecture with E3 equivariance preserves the C3 symmetry and SOC of twisted MoTe2.
    The model's ability to reproduce band degeneracies and Chern numbers relies on this symmetry encoding. Invoked in Sec. II.
  • ad hoc to paper Long-range interactions learned from large-angle twisted structures (9-21 degrees) transfer to small-angle (0.88-5 degrees) structures.
    This is the core transfer learning assumption; the paper provides evidence for angles down to 1.89 degrees, but below that it is unvalidated. Invoked in Sec. II and Supplemental C.
  • domain assumption The sparse PAO Hamiltonian with cutoff Rc is a sufficient representation of the electronic structure for the properties studied.
    All predictions use this Hamiltonian; the accuracy of derived band structures and Chern numbers depends on the representation. Invoked in Sec. II.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transfer learning electronic structure: millielectron volt accuracy for sub-million-atom moir\'e semiconductor." pith.science (2026). https://pith.science/paper/VDQDUVD6

@misc{pith2026250112452,
  author       = {Pith},
  title        = {Pith review of: Transfer learning electronic structure: millielectron volt accuracy for sub-million-atom moir\'e semiconductor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VDQDUVD6}},
  note         = {Machine review of arXiv:2501.12452}
}
abstract

The integration of density functional theory (DFT) with machine learning enables efficient \textit{ab initio} electronic structure calculations for ultra-large systems. In this work, we develop a transfer learning framework tailored for long-wavelength moir\'e systems. To balance efficiency and accuracy, we adopt a two-step transfer learning strategy: (1) the model is pre-trained on a large dataset of computationally inexpensive non-twisted structures until convergence, and (2) the network is then fine-tuned using a small set of computationally expensive twisted structures. Applying this method to twisted MoTe$_2$, the neural network model generates the resulting Hamiltonian for a 1000-atom system in 200 seconds, achieving a mean absolute error below 0.1 meV. To demonstrate $O(N)$ scalability, we model nanoribbon systems with up to 0.25 million atoms ($\sim9$ million orbitals), accurately capturing edge states consistent with predicted Chern numbers. This approach addresses the challenges of accuracy, efficiency, and scalability, offering a viable alternative to conventional DFT and enabling the exploration of electronic topology in large scale moir\'e systems towards simulating realistic device architectures.

Figures

Figures reproduced from arXiv: 2501.12452 by the authors.

Figure 1
Figure 1. FIG. 1. Scheme of transfer learning for electronic structure prediction: A neural network trained solely on a non-twisted [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. DFT-calculated band structures (red curves) and predicted band structures (gray and blue dots) for twisted MoTe [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. (a), (b) Band structures for a twist angle of 3.68 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4. (a) The inference time for the transfer learning neural network model, with the blue bar plot representing the number [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Interference-Enhanced Large Electron-Phonon Coupling from Raman-active Breathing Modes in Moir\'e Semiconductors

    cond-mat.supr-con 2026-07 conditional novelty 7.0 of 10

    An interference selection rule selects reconstruction-matched breathing modes for strong EPC in moiré TMDs, peaking at large twist angles near observed superconductivity.

Reference graph

Works this paper leans on

42 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    and the third band (Chern number 0). Furthermore, for TMD moir´ e systems where SOC is negligible, such as those dominated by Gamma-valley bands at the Fermi level, we can scale up to systems of 0.5 million atoms (twice the size of the SOC case). For simpler twisted graphene or hBN systems, our approach enables scaling to millions of atoms, surpassing the...

    work page

  2. [2]

    Bistritzer and A

    R. Bistritzer and A. H. MacDonald, Proceedings of the National Academy of Sciences 108, 12233 (2011)

    2011

  3. [3]

    J. Cai, E. Anderson, C. Wang, X. Zhang, X. Liu, W. Holtzmann, Y. Zhang, F. Fan, T. Taniguchi, 6 0909 fig3 (a) (b) (c) FIG. 4. (a) The inference time for the transfer learning neural network model, with the blue bar plot representing the number of atoms at twist angle from 5.08 ◦ to 0.88 ◦ and the red line indicating the inference time. The insertion shows...

    work page 2023

  4. [4]

    Y. Zeng, Z. Xia, K. Kang, J. Zhu, P. Kn¨ uppel, C. Vaswani, K. Watanabe, T. Taniguchi, K. F. Mak, and J. Shan, Nature 622, 69 (2023)

    2023

  5. [5]

    F. Xu, Z. Sun, T. Jia, C. Liu, C. Xu, C. Li, Y. Gu, K. Watanabe, T. Taniguchi, B. Tong, et al., Phys. Rev. X 13, 031037 (2023)

    2023

  6. [6]

    H. Park, J. Cai, E. Anderson, Y. Zhang, J. Zhu, X. Liu, C. Wang, W. Holtzmann, C. Hu, Z. Liu, et al., Nature pp. 1–3 (2023)

    work page 2023

  7. [7]

    K. Kang, B. Shen, Y. Qiu, Y. Zeng, Z. Xia, K. Watanabe, T. Taniguchi, J. Shan, and K. F. Mak, Nature 628, 522 (2024)

    2024

  8. [8]

    Y. Cao, V. Fatemi, S. Fang, K. Watanabe, T. Taniguchi, E. Kaxiras, and P. Jarillo-Herrero, Nature 556, 43 (2018)

    2018

Show all 42 references
  1. [9]

    Y. Cao, V. Fatemi, A. Demir, S. Fang, S. L. Tomarken, J. Y. Luo, J. D. Sanchez-Yamagishi, K. Watanabe, T. Taniguchi, E. Kaxiras, et al., Nature 556, 80 (2018)

    2018

  2. [10]

    F. Wu, T. Lovorn, E. Tutuc, and A. H. MacDonald, Phys. Rev. Lett. 121, 026402 (2018), URL https://link.aps. org/doi/10.1103/PhysRevLett.121.026402

    2018 doi

  3. [11]

    Zhang, N

    Y. Zhang, N. F. Q. Yuan, and L. Fu, Phys. Rev. B 102, 201115 (2020), URL https://link.aps.org/doi/ 10.1103/PhysRevB.102.201115

    2020 doi

  4. [12]

    E. C. Regan, D. Wang, C. Jin, M. I. Bakti Utama, B. Gao, X. Wei, S. Zhao, W. Zhao, Z. Zhang, K. Yu- migeta, et al., Nature 579, 359 (2020), URL https: //www.nature.com/articles/s41586-020-2092-4

    2020

  5. [13]

    Y. Tang, L. Li, T. Li, Y. Xu, S. Liu, K. Barmak, K. Watanabe, T. Taniguchi, A. H. MacDonald, J. Shan, et al., Nature 579, 353 (2020), URL https://www. nature.com/articles/s41586-020-2085-3

    2020

  6. [14]

    Wang, E.-M

    L. Wang, E.-M. Shih, A. Ghiotto, L. Xian, D. A. Rhodes, C. Tan, M. Claassen, D. M. Kennes, Y. Bai, B. Kim, et al., Nature materials 19, 861 (2020), URL https:// www.nature.com/articles/s41563-020-0708-6

    2020

  7. [15]

    Ghiotto, E.-M

    A. Ghiotto, E.-M. Shih, G. S. Pereira, D. A. Rhodes, B. Kim, J. Zang, A. J. Millis, K. Watanabe, T. Taniguchi, J. C. Hone, et al., Nature 597, 345 (2021)

    2021

  8. [16]

    T. Li, S. Jiang, L. Li, Y. Zhang, K. Kang, J. Zhu, K. Watanabe, T. Taniguchi, D. Chowdhury, L. Fu, et al., Nature 597, 350 (2021), URL https://www.nature.com/ articles/s41586-021-03853-0

    2021

  9. [17]

    Y. Xu, K. Kang, K. Watanabe, T. Taniguchi, K. F. Mak, and J. Shan (2022), URLhttps://arxiv.org/abs/ 2202.02055

    2022 arXiv

  10. [18]

    Turkel, J

    S. Turkel, J. Swann, Z. Zhu, M. Christos, K. Watanabe, T. Taniguchi, S. Sachdev, M. S. Scheurer, E. Kaxiras, C. R. Dean, et al., Science 376, 193 (2022)

    2022

  11. [19]

    Y. Liu, Y. Gu, T. Bao, N. Mao, C. Li, S. Jiang, L. Liu, D. Guan, Y. Li, H. Zheng, et al., arXiv preprint arXiv:2406.19310 (2024)

    2024

  12. [20]

    Thompson, K

    E. Thompson, K. T. Chu, F. Mesple, X.-W. Zhang, C. Hu, Y. Zhao, H. Park, J. Cai, E. Anderson, K. Watan- abe, et al., arXiv preprint arXiv:2405.19308 (2024)

    2024 arXiv

  13. [21]

    S. Carr, S. Fang, and E. Kaxiras, Nature Reviews Mate- rials 5, 748 (2020)

    2020

  14. [22]

    A. W. Senior, R. Evans, J. Jumper, J. Kirkpatrick, L. Sifre, T. Green, C. Qin, A. ˇZ ´ ıdek, A. W. Nelson, A. Bridgland, et al., Nature 577, 706 (2020)

    2020

  15. [23]

    Zhang, J

    L. Zhang, J. Han, H. Wang, R. Car, and E. Weinan, Phys. Rev. Lett. 120, 143001 (2018)

    2018

  16. [24]

    Behler and M

    J. Behler and M. Parrinello, Physical review letters 98, 146401 (2007)

    2007

  17. [25]

    N. Mao, C. Xu, J. Li, T. Bao, P. Liu, Y. Xu, C. Felser, L. Fu, and Y. Zhang, Communications Physics 7, 262 (2024)

    2024

  18. [26]

    Y. Jia, J. Yu, J. Liu, J. Herzog-Arbeitman, Z. Qi, H. Pi, N. Regnault, H. Weng, B. A. Bernevig, and Q. Wu, Phys- ical Review B 109, 205121 (2024)

    2024

  19. [27]

    Zhang, C

    X.-W. Zhang, C. Wang, X. Liu, Y. Fan, T. Cao, and D. Xiao, Nature Communications 15, 4223 (2024)

    2024

  20. [28]

    H. Li, Z. Wang, N. Zou, M. Ye, R. Xu, X. Gong, W. Duan, and Y. Xu, Nature Computational Science 2, 367 (2022)

    2022

  21. [29]

    X. Gong, H. Li, N. Zou, R. Xu, W. Duan, and Y. Xu, Nature Communications 14, 2848 (2023)

    2023

  22. [30]

    Zhong, H

    Y. Zhong, H. Yu, M. Su, X. Gong, and H. Xiang, npj Computational Materials 9, 182 (2023). 7

    2023

  23. [31]

    Ozaki and H

    T. Ozaki and H. Kino, Phys. Rev. B 69, 195113 (2004)

    2004

  24. [32]

    Weiler, M

    M. Weiler, M. Geiger, M. Welling, W. Boomsma, and T. S. Cohen, Advances in Neural Information Processing Systems 31 (2018)

    2018

  25. [33]

    K. T. Sch¨ utt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. M¨ uller, The Journal of Chem- ical Physics 148 (2018)

    2018

  26. [34]

    S. J. Pan and Q. Yang, IEEE Transactions on knowledge and data engineering 22, 1345 (2009)

    2009

  27. [35]

    T. Bao, R. Xu, H. Li, X. Gong, Z. Tang, J. Fu, W. Duan, and Y. Xu, arXiv preprint arXiv:2404.06449 (2024)

    2024 arXiv

  28. [36]

    F. Wu, T. Lovorn, E. Tutuc, I. Martin, and A. MacDon- ald, Phys. Rev. Lett. 122, 086402 (2019), URL https:// link.aps.org/doi/10.1103/PhysRevLett.122.086402

    2019 doi

  29. [37]

    Kerelsky, L

    A. Kerelsky, L. J. McGilly, D. M. Kennes, L. Xian, M. Yankowitz, S. Chen, K. Watanabe, T. Taniguchi, J. Hone, C. Dean, et al., Nature 572, 95 (2019)

    2019

  30. [38]

    Artaud, L

    A. Artaud, L. Magaud, T. Le Quang, V. Guisset, P. David, C. Chapelier, and J. Coraux, Scientific reports 6, 25670 (2016)

    2016

  31. [39]

    Huder, A

    L. Huder, A. Artaud, T. Le Quang, G. T. De Laissardi` ere, A. G. Jansen, G. Lapertot, C. Chapelier, and V. T. Re- nard, Physical review letters 120, 156405 (2018)

    2018

  32. [40]

    C. Xu, N. Mao, T. Zeng, and Y. Zhang, arXiv preprint arXiv:2403.17003 (2024)

    2024

  33. [41]

    Ozaki, Phys

    T. Ozaki, Phys. Rev. B 67, 155108 (2003)

    2003

  34. [42]

    Morrison, D

    I. Morrison, D. M. Bylander, and L. Kleinman, Phys. Rev. B 47, 6728 (1993). 8 Supplemental Materials A. Details of transfer learning lattice relaxation To achieve lattice relaxation, we employ a two-step transfer learning approach, using the DeePMD-kit code [22]. Initially, we...

    1993

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.