Pith. sign in

REVIEW 4 major objections 4 minor 44 references

A machine-learned tight-binding model computes electronic properties for 100 million atoms, matching graphene's room-temperature mobility.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A machine learning tight-binding framework reconstructs DFT-level Hamiltonians and computes electronic properties for systems with up to 100 million atoms, including graphene mobility versus carrier concentration.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection GPUTB is a credible ML-TB framework with an impressive scale demonstration and a convincing experimental mobility check, but the Hamiltonian is fit only to high-symmetry k-path eigenvalues, leaving off-path DOS and heterojunction transfer claims under-validated, and the missing code/data blocks independent verification. the 4 major comments →

arxiv 2509.06525 v1 pith:5CBJQMZ7 submitted 2025-09-08 cond-mat.mtrl-sci physics.comp-ph

GPUTB: Efficient Machine Learning Tight-Binding Method for Large-Scale Electronic Properties Calculations

Yunlong Wang , Zhixin Liang , Chi Ding , Junjie Wang , Zheyong Fan , Hui-Tian Wang , Dingyu Xing , Jian Sun This is my paper
classification cond-mat.mtrl-sci physics.comp-ph
keywords machine learning tight-bindingGPU accelerationSlater-Kosterlinear-scaling quantum transportelectronic structuregraphene mobilityheterojunctionlarge-scale simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

GPUTB claims that a machine-learned, environment-dependent tight-binding model can deliver density-functional-theory-quality electronic properties at device scale. It trains neural networks to output Slater-Koster hopping integrals and onsite energies from local atomic environments, using DFT band structures as reference, reaching band mean absolute errors of 13-26 meV across tested materials. With those Hamiltonians, a linear-scaling transport solver computes density of states and carrier mobility for systems up to 100 million atoms, including finite-temperature graphene, polycrystalline SiGe, and h-BN/graphene heterojunctions. The central validation is that graphene's room-temperature mobility versus carrier concentration matches experiment, suggesting that device-scale electronic simulations no longer need direct ab-initio calculation of the full system.

Core claim

The paper introduces GPUTB, a machine-learning method that maps an atomic structure directly to a sparse orthogonal tight-binding Hamiltonian. The mapping is trained by matching predicted band structures to DFT reference bands on small periodic cells, with band mean absolute errors reported around 13-26 meV across SiGe, graphene, diamond, GaP, AlAs, and h-BN/graphene. The key claim is that the resulting Hamiltonian is accurate enough beyond the training cell: it is used with a linear-scaling quantum transport solver to compute density of states for a 100-million-atom graphene sheet, room-temperature mobility versus carrier concentration for 6.5-million-atom graphene, and the DOS of million-a

What carries the argument

An environment-dependent Slater-Koster Hamiltonian built from message-passing graph neural networks. Hopping integrals between an atom pair are written as the product of geometric Slater-Koster coefficients and a neural-network output whose input is a Chebyshev-polynomial expansion of the pair's local environment; onsite energies come from a separate network on the atomic environment descriptor. This yields an orthogonal sparse Hamiltonian with no overlap matrix, which is precisely what the linear-scaling transport solver needs to keep every step linear in the number of atoms.

Load-bearing premise

The whole pipeline assumes that a model trained on small perfect cells with finite-temperature snapshots recognizes the same local bonding environments in 100-million-atom, polycrystalline, and heterojunction systems, so the learned parameters transfer without re-fitting and without charge self-consistency.

What would settle it

Take a graphene/h-BN heterojunction with a moiré period not sampled during training, compute its density of states with GPUTB, and compare against a converged DFT band-structure calculation on the same geometry; if the band-gap opening or the DOS near the Fermi level deviates by much more than the claimed 13-26 meV band MAE, the environmental-transfer premise is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Finite-temperature electronic structure and transport become practical for millions of atoms, not just perfect crystals; polycrystalline and heterojunction samples are within reach.
  • A single GPUTB model can describe structurally distinct allotropes and heterostructures, so transfer across phases requires no separate training per structure.
  • The combination with linear-scaling transport means density of states and conductivity are computed in O(N), opening micrometer-scale samples to direct simulation.
  • Graphene's room-temperature mobility as a function of carrier concentration is reproduced from a Hamiltonian trained on small DFT cells, giving a predictive bridge between ab-initio accuracy and experiment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the local-environment descriptor is as transferable as claimed, the same trained model should predict defect states and grain-boundary conductance without retraining; a direct test would be a DFT comparison for a single vacancy in graphene.
  • Because the Hamiltonian is sparse and orthogonal, the sampling of ionic configurations, not the electronic solve, becomes the main remaining cost; combining GPUTB with faster generative sampling would make device-scale workflows routine.
  • The method's fixed basis and finite cutoff mean long-range electrostatic environments, such as metal-water interfaces, are likely outside its reach; testing on such systems would clarify how much expressivity the network needs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. This paper presents GPUTB, a machine-learning tight-binding framework that constructs environment-dependent Slater-Koster Hamiltonians from DFT band structures. The Hamiltonians are used with the linear-scaling quantum transport (LSQT) method to compute density of states, conductivity, and carrier mobility for large systems. Reported band-structure MAEs are 13.0–25.8 meV for SiGe, graphene, diamond, GaP, and AlAs, with transfer from an 8-atom training cell to 64- and 216-atom supercells. The method is applied to graphene DOS with over 100 million atoms, to 2.4–3.0 million atom SiGe single-crystal and polycrystalline systems, and to h-BN/graphene heterojunctions. The room-temperature graphene mobility versus carrier concentration is compared with experimental data. The central claim is that GPUTB provides an accurate and efficient bridge from DFT accuracy to device-scale electronic property calculations.

Significance. If the accuracy and transferability claims are substantiated, GPUTB would be a practically useful tool for device-scale electronic-structure calculations. The strongest evidence is the external experimental benchmark for graphene mobility and the demonstration of DOS calculations on 100-million-atom systems. The paper also shows flexibility with respect to basis sets, exchange-correlation functionals, allotropes, and heterojunctions, and it offers an efficiency comparison against DeePTB. However, the validation has important gaps: training is performed only on high-symmetry k-path eigenvalues, and the large-scale DOS comparisons are made against 0 K DFT tetrahedron DOS, so the claims about off-path Hamiltonian accuracy and finite-temperature DOS are not yet established at the reported level of precision.

major comments (4)
  1. [§2.3, Eq. (3)] The training loss in Eq. (3) is evaluated only on high-symmetry k-path eigenvalues ('High-symmetry k-path selections were made for band structure calculations'). DOS and LSQT conductivity, Eqs. (3)–(4) in §2.4, depend on the full Brillouin zone and on off-path Hamiltonian matrix elements through the velocity operator. Fitting a low-dimensional k-path slice may leave parts of the environment-dependent Hamiltonian underdetermined. No uniform-k-grid validation is reported. I recommend adding a validation set on uniform k-meshes for small cells and reporting band-structure and DOS errors over the full BZ.
  2. [Fig. 3(a), Fig. 4(a)] The finite-temperature DOS for SiGe (Fig. 3a) and for 100-million-atom graphene (Fig. 4a) is compared to a 0 K DFT tetrahedron DOS. This comparison conflates thermal broadening with model error and cannot discriminate between temperature effects and Hamiltonian inaccuracy. A direct comparison of GPUTB DOS against DFT DOS at the same temperature and k-point sampling for small supercells is needed to validate the method quantitatively.
  3. [§3, polycrystalline and heterojunction results] The polycrystalline SiGe DOS/conductivity and the h-BN/graphene heterojunction DOS are presented without quantitative reference DFT calculations for the same disordered or interface geometries. The mid-gap defect states and the heterojunction bandgap opening are discussed only qualitatively. Since GPUTB uses a fixed orthogonal TB Hamiltonian with no charge self-consistency or explicit long-range electrostatics, it is not obvious that these effects are captured reliably. I request benchmarks on small grain-boundary and heterojunction supercells with quantitative error metrics.
  4. [Reproducibility / Data availability] No code or training/test data are provided; the Data Availability statement only says 'Data will be made available on request.' For a machine-learning method paper, this hampers independent verification of the central claims. I encourage the authors to release the code, trained models, and representative training/testing datasets.
minor comments (4)
  1. [Table 1] The table header appears garbled (repeated 'c/c', 'm/c', and 'Average' entries). Please clarify the columns and define how the 'Average' values are computed.
  2. [Equation numbering] Two equations are labeled (3): the conductivity equation in §2.4 and the loss function in §3. Please renumber.
  3. [Fig. 4(c) caption] The caption reads 'Band structure of unit-cell h-BN/graphene heterojunction with 1 million atoms,' which is self-contradictory. Clarify whether the band structure is for a small unit cell and the DOS for the 1-million-atom system.
  4. [Fig. 2 and Fig. 4(b)] MAEs and mobility curves are reported without uncertainties. Please provide error bars or standard deviations over multiple model initializations and/or MD frames to assess statistical significance.

Circularity Check

0 steps flagged

No significant circularity; the large-scale DOS and mobility are derived from a fitted Hamiltonian and benchmarked against independent experiments, not re-imported training targets.

full rationale

The derivation chain is: DFT band eigenvalues (training targets in Eq. 3) -> GPUTB neural-network SK Hamiltonian -> LSQT DOS (Eq. 4) and conductivity -> mobility. The loss function in Eq. 3 contains only band-eigenvalue residuals; neither DOS nor mobility/conductivity appears in any loss term, so the large-scale DOS (100M atoms) and mobility (6.5M atoms) are derived quantities, not fitted inputs renamed as predictions. The mobility-versus-carrier-concentration result is compared with independent experiments (Mayorov et al. [29], Chen et al. [30], Shishir et al. [31]), providing external falsification and breaking any circular loop. The Fig. 4(a) DOS check does reuse the same ABACUS/PBE DFT reference that generated the band-structure training data, but because training is restricted to high-symmetry k-path eigenvalues while DOS integrates over the full Brillouin zone and mobility depends additionally on the velocity operator, the predicted transport properties are not identical to the loss by construction; this is an in-family consistency check, weaker than an independent benchmark, but not a circular reduction. The LSQT engine (Refs [10-13]) and NEP/GPUMD (Refs [28,42,43]) are co-authored tooling, but they are independently published methods, and the experimental agreement here supplies outside support, so these self-citations do not constitute load-bearing circularity. The skeptic's underdetermination concern (path-only training may not fix off-path Hamiltonian matrix elements or the velocity operator) is an accuracy/identifiability risk, not circularity; it means the transport prediction is not entailed by the training data, the opposite of a reduction to inputs. The authors' stated limitation that simple networks may struggle with metal/water interfaces (Sec. 4) concerns scope, not the derivation's self-containedness. No equation-level reduction or fitted-parameter-renamed-as-prediction is present.

Axiom & Free-Parameter Ledger

7 free parameters · 7 axioms · 0 invented entities

The central claim depends on a substantial set of fitted neural network weights and hand-chosen hyperparameters, plus implicit transferability and convergence assumptions. There are no new physical entities. The load-bearing assumptions are that the descriptor representation is complete enough and that small-cell training transfers to device-scale systems.

free parameters (7)
  • cutoff radius r_cut = 5 Å
    Chosen manually to include third nearest neighbors for SiGe and graphene, directly controls Hamiltonian sparsity and accuracy.
  • number of Chebyshev basis functions N_basis
    Descriptor resolution hyperparameter, not specified numerically, affects the featurization quality.
  • MLP weights for onsite, hopping, and environment networks = trained
    All neural network parameters are optimized against DFT band energies, so they are the main fitted parameters of the model.
  • band weights w_b in loss function
    Per-band loss weights are chosen by hand and can bias the fit toward specific bands.
  • finite-temperature training frames = 200 frames per system
    The number of AIMD snapshots used for training is an arbitrary sampling choice that controls how well thermal fluctuations are represented.
  • NEP force-field parameters = RMSE 3.24 meV/atom energy, 89.88 meV/Å force
    Used to generate molecular dynamics structures for large graphene systems; these parameters are fitted to DFT data and their errors propagate into the structures used for validation.
  • iterative band expansion schedule = e.g., band 2 through 26 for SiGe
    The training protocol starts from low bands and adds higher bands step by step; the exact schedule is a modeling choice.
axioms (7)
  • standard math Slater-Koster two-center integral approximation
    The Hamiltonian is built from SK bond integrals between localized orbitals, a standard approximation in tight-binding theory.
  • standard math Orthogonal tight-binding Hamiltonian representation
    The Hamiltonian is assumed orthogonal with no overlap matrix, which is a significant simplification required for the LSQT implementation.
  • domain assumption DFT band structures are reliable reference data
    The model is trained against PBE DFT bands, so the accuracy of GPUTB is bounded by the accuracy of the chosen exchange-correlation functional.
  • ad hoc to paper Chebyshev and message-passing descriptors capture all relevant local chemical environment information
    The paper assumes the descriptor network and environment network are sufficient to predict Hamiltonian elements for unseen atomic environments.
  • ad hoc to paper Small-cell training transfers to large supercells and heterojunctions
    The model trained on 8 to 216 atom cells is applied to 100 million atom systems and h-BN/graphene interfaces without retraining; this transferability is central but is only empirically tested in a few cases.
  • domain assumption LSQT with 200 propagation steps and random states converges
    The paper assumes the Chebyshev expansion and velocity autocorrelation calculation are converged after 200 steps for all reported systems.
  • domain assumption NEP force field structures are representative of true finite-temperature configurations
    For large graphene MD, the structures come from a machine-learned NEP potential rather than DFT, so the quality of the finite-temperature test depends on the accuracy of that potential.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GPUTB: Efficient Machine Learning Tight-Binding Method for Large-Scale Electronic Properties Calculations." pith.science (2026). https://pith.science/paper/5CBJQMZ7

@misc{pith2026250906525,
  author       = {Pith},
  title        = {Pith review of: GPUTB: Efficient Machine Learning Tight-Binding Method for Large-Scale Electronic Properties Calculations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5CBJQMZ7}},
  note         = {Machine review of arXiv:2509.06525}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The high computational cost of ab-initio methods limits their application in predicting electronic properties at the device scale. Therefore, an efficient method is needed to map the atomic structure to the electronic structure quickly. Here, we develop GPUTB, a GPU-accelerated tight-binding (TB) machine learning framework. GPUTB employs atomic environment descriptors, enabling the model parameters to incorporate environmental dependence. This allows the model to transfer to different basis, xc-functionals, and allotropes easily. Combined with the linear scaling quantum transport method, we have calculated the electronic density of states for up to 100 million atoms in pristine graphene. Trained on finite-temperature structures, the model can be easily extended to millions of atom finite-temperature systems. Furthermore, GPUTB can also successfully describe h-BN/graphene heterojunction systems, demonstrating its capability to handle complex material with high precision. We accurately reproduce the relationship between carrier concentration and room temperature mobility in graphene to verify the framework's accuracy. Therefore, our GPUTB framework presents a delicate balance between computational accuracy and efficiency, providing a powerful computational tool for investing electronic properties for large systems with millions of atoms.

Figures

Figures reproduced from arXiv: 2509.06525 by Chi Ding, Dingyu Xing, Hui-Tian Wang, Jian Sun, Junjie Wang, Yunlong Wang, Zheyong Fan, Zhixin Liang.

Figure 1
Figure 1. Figure 1: Flowchart and network architecture of GPUTB calculations. (a) The embedding network of GPUTB, where vi, vj is the node message, eij is the edge message, Zi, Zj indicates the types of atoms, and rij, rik, rjk is the distance between neighboring atom pairs ij, ik, jk, respectively. (b) Descriptor network and residual parameter network, MLP represents Multi￾Layer Perceptron, whereas the Hamiltonian components… view at source ↗
Figure 2
Figure 2. Figure 2: Generalization and accuracy of the model at different scales. (a-c) The error histogram of bands by model trained with 8 atoms predicted for 8 atoms, 64 atoms (2×2×2 supercell), and 216 atoms (3×3×3 supercell) systems compared to the corresponding DFT energies, and the inset shows the energy error distribution. (d) Comparison of the time cost of calculating Γ-point eigenvalues with GPUTB, DFT-PW, and DFT-L… view at source ↗
Figure 3
Figure 3. Figure 3: Calculate the DOS and conductivity for single-crystal and polycrystalline SiGe at finite temperatures. (a) Comparison of finite-temperature DOS for SiGe at 300K with DFT￾calculated DOS for the unit cell at 0 K. (b) Comparison of finite-temperature DOS for single-crystal and polycrystalline SiGe. (c) Comparison of finite-temperature conductivity for single-crystal and polycrystalline SiGe. (d) Evolution of … view at source ↗
Figure 4
Figure 4. Figure 4: Device-scale graphene DOS, carrier mobility, and the DOS of h-BN/graphene heterojunctions. (a) Comparison of the DOS of graphene containing 100 million atoms at finite-temperature with the DOS computed using the DFT tetrahedron method for a graphene unit cell. (b) Carrier mobility of graphene at 300 K as a function of carrier concentration, compared with previous experiments (Mayorov et al. [29] and Chen e… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 1 internal anchor

  1. [1]

    L. E. Ratcliff, S. Mohr, G. Huhs, T. Deutsch, M. Masella, and L. Genovese, Challenges in large scale quantum mechanical calculations, WIREs Comput Mol Sci 7, e1290 (2017)

    work page 2017

  2. [2]

    Galli and M

    G. Galli and M. Parrinello, Large scale electronic structure calculations, Phys. Rev. Lett. 69, 3547 (1992)

    work page 1992

  3. [3]

    Ozaki and H

    T. Ozaki and H. Kino, Efficient projector expansion for the ab initio LCAO method, Phys. Rev. B 72, 045121 (2005)

    work page 2005

  4. [4]

    Sánchez-Portal, P

    D. Sánchez-Portal, P. Ordejón, E. Artacho, and J. M. Soler, Density-functional method for very large systems with LCAO basis sets, International Journal of Quantum Chemistry 65, 453 (1997)

    work page 1997

  5. [5]

    Zhang, B

    X. Zhang, B. Grabowski, F. Körmann, C. Freysoldt, and J. Neugebauer, Accurate electronic free energies of the 3d,4d, and 5d transition metals at high temperatures, Phys. Rev. B 95, (2017)

    work page 2017

  6. [6]

    Ozaki and H

    T. Ozaki and H. Kino, Numerical atomic basis orbitals from H to Kr, Phys. Rev. B 69 , 195113 (2004)

    work page 2004

  7. [7]

    M. Chen, G. -C. Guo, and L. He, Systematically improvable optimized atomic basis sets for ab initio calculations, J. Phys.: Condens. Matter 22, 445501 (2010)

    work page 2010

  8. [8]

    P. Li, X. Liu, M. Chen, P. Lin, X. Ren, L. Lin, C. Yang, and L. He, Large -scale ab initio simulations based on systematically improvable atomic basis, Computational Materials Science 112, 503 (2016)

    work page 2016

  9. [9]

    Roche and D

    S. Roche and D. Mayou, Conductivity of Quasiperiodic Systems: A Numerical Study, Phys. Rev. Lett. 79, 2518 (1997)

    work page 1997

  10. [10]

    Z. Fan, A. Uppstu, T. Siro, and A. Harju, Efficient linear -scaling quantum transport calculations on graphics processing units and applications on electron transport in graphene, Computer Physics Communications 185, 28 (2014)

    work page 2014

  11. [11]

    Z. Fan, V . Vierimaa, and A. Harju, GPUQT: An efficient linear-scaling quantum transport code fully implemented on graphics processing units, Computer Physics Communications 230, 113 (2018)

    work page 2018

  12. [12]

    Z. Fan, J. H. Garcia, A. W. Cummings, J. E. Barrios -Vargas, M. Panhans, A. Harju, F. Ortmann, and S. Roche, Linear scaling quantum transport methodologies, Physics Reports 903, 1 (2021)

    work page 2021

  13. [13]

    Z. Fan, Y . Xiao, Y . Wang, P. Ying, S. Chen, and H. Dong, Combining linear -scaling quantum transport and machine -learning molecular dynamics to study thermal and electronic transports in complex materials, J. Phys.: Condens. Matter 36, 245901 (2024)

    work page 2024

  14. [14]

    Hu et al., Observation of Rydberg moiré excitons, Science 380, 1367 (2023)

    Q. Hu et al., Observation of Rydberg moiré excitons, Science 380, 1367 (2023)

    work page 2023

  15. [15]

    S. Cui, C. Jiang, Z. Zhan, T. Wilson, N. Zhang, X. Xie, S. Yuan, H. Wang, C. Lewandowski, and G. Ni, Nanoscale Optical Conductivity Imaging of Double -Moiré Twisted Bilayer Graphene, Nano Lett. 24, 11490 (2024)

    work page 2024

  16. [16]

    V ogl, H

    P. V ogl, H. P. Hjalmarson, and J. D. Dow, A Semi-empirical tight-binding theory of the electronic structure of semiconductors†, Journal of Physics and Chemistry of Solids 44, 365 (1983)

    work page 1983

  17. [17]

    Jancu, R

    J.-M. Jancu, R. Scholz, F. Beltram, and F. Bassani, Empirical spds * tight-binding 15 / 21 calculation for cubic semiconductors: General method and material parameters, Phys. Rev. B 57, 6493 (1998)

    work page 1998

  18. [18]

    Nakhaee, S

    M. Nakhaee, S. A. Ketabi, and F. M. Peeters, Tight-Binding Studio: A technical software package to find the parameters of tight -binding Hamiltonian, Computer Physics Communications 254, 107379 (2020)

    work page 2020

  19. [19]

    Aradi, B

    B. Aradi, B. Hourahine, and Th. Frauenheim, DFTB+, a Sparse Matrix -Based Implementation of the DFTB Method, J. Phys. Chem. A 111, 5678 (2007)

    work page 2007

  20. [20]

    M. Gaus, Q. Cui, and M. Elstner, DFTB3: Extension of the Self -Consistent-Charge Density-Functional Tight-Binding Method (SCC -DFTB), J. Chem. Theory Comput. 7 , 931 (2011)

    work page 2011

  21. [21]

    Learning local equivariant representations for quantum operators

    Z. Zhouyin, Z. Gan, S. K. Pandey, L. Zhang, and Q. Gu, Learning Local Equivariant Representations for Quantum Operators, arXiv:2407.06053

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    X. Gong, H. Li, N. Zou, R. Xu, W. Duan, and Y . Xu, General framework for E(3) - equivariant neural network representation of density functional theory Hamiltonian, Nat Commun 14, 2848 (2023)

    work page 2023

  23. [23]

    Zhong, H

    Y . Zhong, H. Yu, M. Su, X. Gong, and H. Xiang, Transferable equivariant graph neural networks for the Hamiltonians of molecules and solids, Npj Comput Mater 9, 182 (2023)

    work page 2023

  24. [24]

    A. A. Mostofi, J. R. Yates, Y .-S. Lee, I. Souza, D. Vanderbilt, and N. Marzari, wannier90: A tool for obtaining maximally -localised Wannier functions, Computer Physics Communications 178, 685 (2008)

    work page 2008

  25. [25]

    A. A. Mostofi, J. R. Yates, G. Pizzi, Y .-S. Lee, I. Souza, D. Vanderbilt, and N. Marzari, An updated version of wannier90: A tool for obtaining maximally-localised Wannier functions, Computer Physics Communications 185, 2309 (2014)

    work page 2014

  26. [26]

    Q. Gu, L. Zhang, and J. Feng, Neural network representation of electronic structure from ab initio molecular dynamics, Science Bulletin 67, 29 (2022)

    work page 2022

  27. [27]

    Q. Gu, Z. Zhouyin, S. K. Pandey, P. Zhang, L. Zhang, and W. E, Deep learning tight- binding approach for large -scale electronic simulations at finite temperatures with ab initio accuracy, Nat Commun 15, 6772 (2024)

    work page 2024

  28. [28]

    Z. Fan, Z. Zeng, C. Zhang, Y . Wang, K. Song, H. Dong, Y . Chen, and T. Ala -Nissila, Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport, Phys. Rev. B 104, 104309 (2021)

    work page 2021

  29. [29]

    A. S. Mayorov et al., Micrometer -Scale Ballistic Transport in Encapsulated Graphene at Room Temperature, Nano Lett. 11, 2396 (2011)

    work page 2011

  30. [30]

    J.-H. Chen, C. Jang, S. Xiao, M. Ishigami, and M. S. Fuhrer, Intrinsic and extrinsic performance limits of graphene devices on SiO2, Nature Nanotech 3, 206 (2008)

    work page 2008

  31. [31]

    R. S. Shishir and D. K. Ferry, Intrinsic mobility in graphene, J. Phys.: Condens. Matter 21, 232204 (2009)

    work page 2009

  32. [32]

    M. Gori, G. Monfardini, and F. Scarselli, A New Model for Learning in Graph Domains , in Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005. , V ol. 2 (IEEE, Montreal, Que., Canada, 2005), pp. 729–734

    work page 2005

  33. [33]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, Neural Message Passing for Quantum Chemistry, in Proceedings of the 34th International Conference on Machine Learning, edited by D. Precup and Y . W. Teh, V ol. 70 (PMLR, 2017), pp. 1263–1272

    work page 2017

  34. [34]

    J. C. Slater and G. F. Koster, Simplified LCAO Method for the Periodic Potential Problem, 16 / 21 Phys. Rev. 94, 1498 (1954)

    work page 1954

  35. [35]

    [9-12]., (n.d.)

    See Supplemental Material [url] for Training and Electronic Structure Calculation Methods, which includes Refs. [9-12]., (n.d.)

    work page

  36. [36]

    Ben Mahmoud, A

    C. Ben Mahmoud, A. Anelli, G. Csányi, and M. Ceriotti, Learning the electronic density of states in condensed matter, Phys. Rev. B 102, 235130 (2020)

    work page 2020

  37. [37]

    Z.-G. Chen, Z. Shi, W. Yang, X. Lu, Y . Lai, H. Yan, F. Wang, G. Zhang, and Z. Li, Observation of an intrinsic bandgap and Landau level renormalization in graphene/boron - nitride heterostructures, Nat Commun 5, 4461 (2014)

    work page 2014

  38. [38]

    J. Jung, A. M. DaSilva, A. H. MacDonald, and S. Adam, Origin of band gaps in graphene on hexagonal boron nitride, Nat Commun 6, 6308 (2015)

    work page 2015

  39. [39]

    Yankowitz, J

    M. Yankowitz, J. Xue, D. Cormode, J. D. Sanchez-Yamagishi, K. Watanabe, T. Taniguchi, P. Jarillo-Herrero, P. Jacquod, and B. J. LeRoy, Emergence of superlattice Dirac points in graphene on hexagonal boron nitride, Nature Phys 8, 382 (2012)

    work page 2012

  40. [40]

    P. Lin, X. Ren, X. Liu, and L. He, Ab initio electronic structure calculations based on numerical atomic orbitals: Basic fomalisms and recent progresses, WIREs Comput Mol Sci 14, e1687 (2024)

    work page 2024

  41. [41]

    P. Lin, X. Ren, and L. He, Strategy for constructing compact numerical atomic orbital basis sets by incorporating the gradients of reference wavefunctions, Phys. Rev. B 103 , 235131 (2021)

    work page 2021

  42. [42]

    Z. Fan, W. Chen, V . Vierimaa, and A. Harju, Efficient molecular dynamics simulations with many-body potentials on graphics processing units, Computer Physics Communications 218, 10 (2017)

    work page 2017

  43. [43]

    Song et al., General -purpose machine-learned potential for 16 elemental metals and their alloys, Nat Commun 15, 10208 (2024)

    K. Song et al., General -purpose machine-learned potential for 16 elemental metals and their alloys, Nat Commun 15, 10208 (2024)

    work page 2024

  44. [44]

    Schlipf and F

    M. Schlipf and F. Gygi, Optimization algorithm for the generation of ONCV pseudopotentials, Computer Physics Communications 196, 36 (2015). 17 / 21 Figures captions Fig. 1. Flowchart and network architecture of GPUTB calculations. (a) The embedding network of GPUTB, where vi, vj is the node message, e ij is the edge message, Zi, Zj indicates the types of ...

    work page 2015

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.