Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

A machine-learning model trained on small displaced bilayers predicts electron densities of large twisted moiré superlattices, provided its descriptors encode long-range electrostatics.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 02:37 UTC pith:Y5NJXBZJ

load-bearing objection Worth reading, but the headline accuracy numbers are tuned on the test set, so treat them as optimistic until an honest retest. the 4 major comments →

arxiv 2602.09938 v2 pith:Y5NJXBZJ submitted 2026-02-10 cond-mat.mtrl-sci physics.chem-phphysics.comp-ph

Long-Range Machine Learning of Electron Density for Twisted Bilayer Moir\'e Materials

Zekun Lou , Alan M. Lewis , Mariana Rossi This is my paper
classification cond-mat.mtrl-sci physics.chem-phphysics.comp-ph
keywords moiré superlatticeselectron density predictionmachine learningGaussian process regressionlong-range descriptorstwisted bilayerband structurespin-orbit coupling
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.

This paper contends that the nearsightedness/locality assumption built into most machine-learning electronic-structure models breaks down for moiré superlattices, where interlayer polarization and charge rearrangement extend over nanometres. The authors train a Gaussian process model (SALTED) on electron densities of small 3×3 displaced bilayer cells only, then use it to predict densities of large twisted bilayer supercells with thousands of atoms. They find that a purely long-range descriptor, LOVV (built from the tensor product of Coulomb-like atomic electrostatic representations), is necessary for reliable extrapolation: it keeps low-energy band-structure errors below about 5 meV across five materials (graphene, hBN, TiS2, ZrS2, MoS2) at superlattice sizes up to ~50 Å, whereas local (SOAP) and mixed-range (LODE) descriptors degrade or fail for several of these systems. The same densities support subsequent derived properties—flat-band narrowing, spin-orbit splittings, and domain-wall electric fields—making first-principles-quality studies of small-angle moiré physics computationally feasible.

Core claim

In the paper's own terms, the discovery is that SALTED, trained only on small displaced bilayers, can extrapolate electron densities to large twisted moiré structures if the atomic environment is represented by the LOVV descriptor, whose |V⊗V⟩ structure encodes correlations between long-range electrostatic fields. This long-range representation keeps the descriptors of twisted test structures inside the training manifold, allowing Gaussian process regression to interpolate rather than extrapolate in descriptor space. The paper also shows that global density error is not a reliable guide to band-structure accuracy: for MoS2, a descriptor with the lowest density error (SOAP) produced catastrop

What carries the argument

The load-bearing object is the LOVV descriptor, defined as the tensor product of atomic electrostatic representations (V⊗V), where each atom's V is a Coulomb-like 1/r potential coming from a smeared atomic density. This gives the descriptor an effective sensitivity range of about 17 Å, compared to 6 Å for the local SOAP descriptor and 13 Å for the mixed LODE descriptor. The paper's argument is that only such a purely long-range representation captures the nanometre-scale interlayer charge rearrangement and polarization that govern moiré electronic structure, and—equally important—places twisted-bilayer test structures within the descriptor-space distribution of the displaced-bilayer training

Load-bearing premise

The central assumption is that the descriptors of twisted and relaxed test structures lie within the descriptor-space manifold spanned by the displaced-bilayer training set, so that the Gaussian process model is interpolating; the paper's evidence for this is a qualitative 2-D projection, and the smallest-angle extrapolations (beyond ~50 Å, up to 4000+ atoms) have no DFT reference.

What would settle it

A DFT band-structure calculation for a relaxed twisted bilayer at θ ≈ 2° or below (e.g., TiS2 or hBN) that disagrees with the SALTED-LOVV prediction by more than about 10 meV in the low-energy bands, or a descriptor-space analysis in the full (non-projected) descriptor space showing test structures outside the training distribution, would falsify the extrapolation claim.

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

If this is right

  • Systematic ab initio-type band-structure maps as a function of twist angle become feasible for supercells of thousands of atoms, including angles below 3° where DFT is prohibitive.
  • Spin-orbit coupling and other non-self-consistent perturbations can be added to predicted densities without retraining, opening valleytronic and topological studies of twisted TMDCs.
  • Real-space electrostatic observables such as domain-wall electric fields in relaxed twisted hBN can be computed directly from predicted densities, linking ML density prediction to ferroelectric phenomena.
  • Bandwidth narrowing and possible gap closure at small twist angles (e.g., predicted gap closure in TiS2 below ~3°) are concrete, testable predictions that experiments or future DFT benchmarks could check.

Where Pith is reading between the lines

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

  • If the descriptor-manifold explanation is right, the same strategy of training on displaced bilayers with long-range descriptors should transfer to other layered systems, including twisted trilayers or heterobilayers with lattice mismatch, without retraining on large cells.
  • The paper's analysis suggests that other purely long-range descriptor constructions (beyond LOVV) might also work, and that a practical design rule is to optimize descriptors and regularization for downstream band-structure error rather than global density error.
  • The predicted flat-band and gap-closure trends at sub-3° twist angles are falsifiable by experiment (e.g., transport or ARPES) once such samples become available, or by a future DFT calculation at those sizes.

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 / 3 minor

Summary. The paper presents an extension of the SALTED Gaussian-process-regression framework for predicting electron densities of twisted bilayer moiré materials. The model is trained exclusively on small (3×3) displaced bilayer configurations and then applied to much larger twisted supercells (up to ~1000–5000 atoms). Three descriptors are compared: local SOAP, long-range LODE, and a purely long-range electrostatic descriptor LOVV. The authors report that LOVV provides robust extrapolation, with low-energy band-structure errors ≲5 meV for five materials (graphene, hBN, TiS2, ZrS2, MoS2), and demonstrate applications to band-width narrowing, spin-orbit coupling, relaxed magic-angle twisted bilayer graphene, and electric fields at hBN domain walls. The central claim is that long-range descriptors are necessary for density-based ML to capture moiré-scale electronic structure.

Significance. If the central claim holds, this is a meaningful advance: it shows that a density-based ML method can go beyond local descriptors and access moiré-scale electronic structure directly from small-cell training data, with direct DFT validation up to ~1300 atoms. The paper is careful to evaluate downstream band structures rather than only density metrics, and it makes concrete, falsifiable predictions for flat bands, SOC splittings, and real-space electric fields. The public data-availability plan and the descriptor sensitivity analysis are also strengths. However, the current evaluation protocol does not provide a clean out-of-sample measurement of the headline accuracy, and several claims are broader than the evidence.

major comments (4)
  1. [Section IV E / Fig. 6 / SI Table IV] The headline errors are not clean out-of-sample. Fig. 6 and SI Table IV show that the GPR regularization η (and implicitly the descriptor choice) is selected by minimizing the Moiré Band MAE on the very same twisted structures that are used to report the extrapolation errors. With only 10–16 twisted geometries per material, tuning one scalar η and selecting among SOAP/LODE/LOVV on the evaluation set can systematically lower the reported numbers. This affects Fig. 1(A) and the abstract's '<5 meV' claim. Please provide an a priori evaluation, for example leave-one-angle-out cross-validation or a separate held-out set of twist angles, and report both the selected and the a priori errors. At minimum, the paper should state explicitly that the reported numbers are post-selection and discuss the likely bias.
  2. [Section II A vs Section II D / Fig. 4] The claim 'only LOVV maintains robust extrapolation with errors ≲5 meV across all materials and system sizes exceeding 1000 atoms' is inconsistent with the magic-angle TBG result in Fig. 4(A), where the SOAP-based model gives a mean absolute error of about 15 meV for a 11908-atom cell. If the 5 meV claim is intended only for LOVV and/or only for systems up to ~50 Å (as in the DFT-validated part of Fig. 1), this should be stated explicitly. As written, the scope of the claim is broader than the evidence.
  3. [Section III and SI S4, Fig. 31] The speedup claim is inconsistent. The main text says 'LOVV-based SALTED models remain 10 to 100 times faster than fully converged DFT,' and the introduction says 'speedup between one and two orders of magnitude.' However, SI S4 reports that for a 1000-atom system, LOVV-based SALTED is about 4× faster for hBN and about 20× faster for ZrS2, and even doubling these values gives 8× and 40×. The paper should either revise the claimed speedups or provide benchmarks that support them.
  4. [Section IV E / SI S5B / Fig. 2] The extrapolation to 3768–4564 atoms (Fig. 2) has no DFT reference, and the only support for descriptor-space coverage in that regime is the UMAP analysis. The SI itself acknowledges that UMAP is qualitative and can produce density artifacts. Please add a quantitative coverage measure (e.g., distances in descriptor space from training data, or prediction uncertainties) or explicitly temper claims for twist angles below the validated range. As it stands, the 3768–4564-atom predictions rest on an assumption that is acknowledged but not quantitatively tested.
minor comments (3)
  1. [SI S1C / Fig. 9 caption] There is an inconsistency in the reported k-path sampling: SI S1C says '11 points on KΓ, 9 points on ΓM, and 6 points on MK', while the caption of Fig. 9 says '11 points on KΓ, 6 points on ΓM, and 9 points on MK'. Please correct.
  2. [Table I caption vs Figure 1] Table I states that hyperparameters are optimized for density prediction accuracy, but Figure 1 uses models selected for band-structure accuracy (Section IV F). Clarify which models are used for which reported numbers.
  3. [Throughout] Minor terminology/typos: 'deregistration' is used where 'relaxation' or 'atomic displacement' is meant; figure panel labels mix uppercase and lowercase (e.g., Fig. 5 uses '(c)' while others use '(A)'). Please proofread.

Circularity Check

1 steps flagged

Reported moiré band errors are partially in-sample: GPR regularization (and per-material descriptor choice) is selected on the same twisted-bilayer test set whose band MAE is then reported as the extrapolation error.

specific steps
  1. fitted input called prediction [Section II A (Fig. 1A), Section IV E (Fig. 6), SI Table IV]
    "SALTED models used here are those with the best band structure prediction performance for each material and descriptor, obtained by following the model optimisation workflow, see Section IV F and Supporting Information S2. ... However, the error in the band structure prediction of the twisted bilayers shows contrasting dependence on η across material classes, as shown in Figure 6. ... Recommended hyperparameters for moiré band structure prediction for different descriptors and materials following the SALTED framework guidelines."

    The headline 'only LOVV maintains robust extrapolation with errors ≲5 meV' is quantified by Moiré Band MAE (Eq. 8). That same metric is used to select the GPR regularization η (Fig. 6) and to choose the best descriptor/model per material (Fig. 1 caption; SI Table IV). The reported extrapolation error is therefore the minimized selection objective on the twisted-bilayer test structures, not an independent out-of-sample measurement. The density training set remains displaced 3×3 bilayers, so the circularity is partial: it taints the quantitative accuracy claim and the LOVV-vs-LODE/SOAP comparison, while the existence of a density-based extrapolation is not itself manufactured.

full rationale

The paper's core derivation—training SALTED on displaced-bilayer densities and using the predicted density to reconstruct band structures of twisted bilayers—is not circular in the sense of using twisted densities as training labels. The DFT band structures used for validation are independent of the density training set, and the LOVV descriptor is taken from prior literature rather than invented here. However, the central quantitative claim is weakened by test-set leakage: Figure 6 and SI Table IV show that η, and effectively the descriptor choice, are optimized against Moiré Band MAE on the same twisted structures whose errors are then reported in Figure 1A as extrapolation accuracy. Thus the ≤5 meV figure and the 'only LOVV' conclusion are partly in-sample. This is a model-selection/validation flaw rather than a full equivalence-by-construction, so it does not warrant a score of 6 or higher. No load-bearing self-citation chain, imported uniqueness theorem, or renaming of a known result was found; prior SALTED self-citations are methodological foundations, not the claimed result.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

The central claim rests on per-material GPR hyperparameters, the density-fitting/low-rank representation, and the assumption that descriptor-space overlap guarantees extrapolation. No new physical entities are introduced; LOVV is a descriptor from the prior literature.

free parameters (5)
  • GPR regularisation η = per material/descriptor: 1e-14 to 1e-3 (Tables IV, V)
    Optimized per material and descriptor; for band prediction it is selected using moiré test band-error curves, so the reported accuracy depends directly on these chosen values.
  • Singular-value truncation threshold δ = 0 (graphene, hBN); 1e-6 (TiS2); 1e-2 (ZrS2, MoS2)
    Chosen empirically to stabilize ill-conditioned density-fitting overlap matrices; affects the TMDC models on which the LOVV claims rest.
  • Descriptor hyperparameters = rcut=6 Å, n_rad=6, n_ang=6, σ=0.3 Å, z=2, M_env=200
    Fixed globally across materials and descriptors; central to the SOAP/LODE/LOVV comparison.
  • MACE MLIP hyperparameters = TBG: hidden 32x0e+32x1o, 2 interactions; hBN: 64x0e+64x1o+64x2e
    Used to relax TBG and TB-hBN structures; force MAEs are reported, but relaxation accuracy enters the electronic-structure predictions in the case studies.
  • Domain-wall saturation fit (w_eq, a0) = w_eq=4.23±0.01 nm, a0=15.9±0.1 nm
    Fitted to relaxed TB-hBN domain-wall widths; used to interpret the electric-field saturation, though not part of the central ML claim.
axioms (7)
  • domain assumption PBE DFT in FHI-aims provides accurate reference densities and band structures for all training and validation data.
    Every ML target and every validation band structure is PBE-DFT; inaccuracies of PBE for correlated moiré physics are inherited by the method.
  • standard math The electron density uniquely determines ground-state properties, so a one-shot non-self-consistent DFT restart from a predicted density is a meaningful evaluation.
    Invoked through the Hohenberg-Kohn theorem (Ref. 39) and the half-SCF procedure in Sec. IV D.
  • domain assumption The density-fitting expansion with the chosen auxiliary basis and singular-value truncation preserves the physically relevant parts of the density.
    The low-rank approximation in Sec. IV C discards near-null-space directions; if this truncation removes relevant physics, downstream properties are affected.
  • domain assumption Descriptor-space overlap between training and test structures is sufficient for GPR extrapolation.
    Assumed throughout Sec. IV E and supported only qualitatively by UMAP; the paper acknowledges UMAP's limitations.
  • domain assumption Commensurate twist angles generated by Eq. S11 with r=1 represent the moiré physics of interest.
    All twisted test structures use this integer parametrization; it covers selected commensurate angles, not a continuous distribution.
  • domain assumption MACE machine-learning interatomic potentials trained on PBE+MBD data produce relaxed geometries accurate enough for electronic-structure prediction.
    Used for TBG and TB-hBN relaxation case studies; errors in relaxed geometries propagate into SALTED predictions.
  • standard math Continuum elasticity formula for soliton domain-wall width (Eq. S32) is applicable to relaxed TB-hBN.
    Used to interpret the measured domain-wall saturation; the paper compares it with its own fitted value.

pith-pipeline@v1.3.0-alltime-deepseek · 39148 in / 13622 out tokens · 136659 ms · 2026-08-03T02:37:16.528591+00:00 · methodology

0 comments
read the original abstract

Moir\'e superlattices in two-dimensional (2D) materials exhibit rich quantum phenomena, but ab initio modelling of these systems remains computationally prohibitive. Existing machine learning methods for accelerating density-functional theory (DFT) can target the prediction of different quantities and often rely on the locality assumption. Here we train a Gaussian process regression SALTED model exclusively on the electron densities of small displaced bilayer structures and then extrapolate electron density prediction to the large supercells required to describe small twist angles between these bilayers. We show the necessity of long-range descriptors to yield reliable band structures and electrostatic properties of large twisted bilayer structures, when these are derived from predicted densities. We demonstrate that the choice of descriptor determines the distribution of residual density errors, which in turn affects the downstream electronic properties. We apply our models to twisted bilayer graphene, hexagonal boron nitride, and transition metal dichalcogenides, focusing on the model's capacity to predict complex phenomena, including flat band formation, bandwidth narrowing, domain-wall electric fields, and spin-orbit coupling effects. Beyond moir\'e materials, this approach provides a general methodology for electronic structure prediction in large-scale systems with substantial long-range phenomena related to non-local geometric information.

Figures

Figures reproduced from arXiv: 2602.09938 by Alan M. Lewis, Mariana Rossi, Zekun Lou.

Figure 1
Figure 1. Figure 1: (A) illustrates the dependence of the accuracy of the extrapolated band structures on both the choice of descriptor and the system size. In [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: shows the distribution of the singular val￾ues of the overlap matrices for every structure in the training dataset of each material, along with the aver￾age condition numbers of these matrices. The condition numbers indicate particularly severe ill-conditioning for the TMDC training datasets, likely due to the larger and more diffuse DF auxiliary basis sets required by heavier elements. As a result, the lo… view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10 [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: FIG. 12 [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: FIG. 13 [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: FIG. 14 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: FIG. 15 [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: FIG. 16 [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: FIG. 17 [PITH_FULL_IMAGE:figures/full_fig_p027_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: FIG. 18 [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: FIG. 19 [PITH_FULL_IMAGE:figures/full_fig_p029_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: FIG. 20 [PITH_FULL_IMAGE:figures/full_fig_p030_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: FIG. 21 [PITH_FULL_IMAGE:figures/full_fig_p031_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: FIG. 22 [PITH_FULL_IMAGE:figures/full_fig_p032_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Relaxed structures, with their spatially varying interlayer distances and local atomic displacements, natu￾rally locate within this high-dimensional manifold of the training dataset. In contrast, rigid twisted bilayers that have uniform interlayer distances form a low-dimensional submanifold that is under-represented in the training set. Since GPR prediction accuracy depends on descriptor overlap between … view at source ↗
Figure 24
Figure 24. Figure 24: FIG. 24 [PITH_FULL_IMAGE:figures/full_fig_p034_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: FIG. 25 [PITH_FULL_IMAGE:figures/full_fig_p034_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: FIG. 26 [PITH_FULL_IMAGE:figures/full_fig_p035_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: demonstrates the spatial error distributions of density and Hartree potential. While the density prediction error is not significant, the Hartree potential exhibits an AA-localised deviation. Despite this pattern, the in-plane electric field at the AB-BA domain wall is not significantly affected due to cancellation between symmetric AA-localised deviation pairs on both sides of the domain wall centre. (a)… view at source ↗
Figure 28
Figure 28. Figure 28: FIG. 28 [PITH_FULL_IMAGE:figures/full_fig_p037_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: FIG. 29 [PITH_FULL_IMAGE:figures/full_fig_p038_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: FIG. 30 [PITH_FULL_IMAGE:figures/full_fig_p039_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: FIG. 31 [PITH_FULL_IMAGE:figures/full_fig_p041_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: shows the distance-dependent sensitivity for all three descriptors. SOAP is limited by its cutoff radius rcut = 6.0 ˚A, confirming its locality character. LODE and LOVV maintain sensitivity at extended distance, and have identical decay behaviour for r > rcut. This similarity reflects their shared electrostatic 1/r character in the long￾range regime. We interpolate the log-log plot linearly to obtain the … view at source ↗
Figure 33
Figure 33. Figure 33: demonstrates this artifact using synthetic Gaussian data. One can observe that, although the prediction set distribution is fully contained within the training set in high-dimensional space, the 2D UMAP projection shows increasing separation as the prediction set density increases. To compensate for this artifact and enable fair comparison across materials and descriptors, we apply density-based subsampli… view at source ↗
Figure 34
Figure 34. Figure 34: FIG. 34 [PITH_FULL_IMAGE:figures/full_fig_p045_34.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Reconstructing local environments from concise atomistic representations

    physics.comp-ph 2026-07 conditional novelty 6.0

    Atomic environments can be recovered from what amounts to dozens of rotation-invariant numbers, and the same inversion reveals new pairs of distinct geometries that the descriptors cannot tell apart.

Reference graph

Works this paper leans on

95 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    Y. Cao, V. Fatemi, S. Fang, K. Watanabe, T. Taniguchi, E. Kaxiras, and P. Jarillo-Herrero, Nature556, 43 (2018)

    2018

  2. [2]

    L. Xian, D. M. Kennes, N. Tancogne-Dejean, M. Altarelli, and A. Rubio, Nano Letters19, 4934 (2019)

    2019

  3. [3]

    A. L. Sharpe, E. J. Fox, A. W. Barnard, J. Finney, K. Watanabe, T. Taniguchi, M. A. Kastner, and D. Goldhaber-Gordon, Science365, 605 (2019)

    2019

  4. [4]

    Yasuda, X

    K. Yasuda, X. Wang, K. Watanabe, T. Taniguchi, and P. Jarillo-Herrero, Science372, 1458 (2021)

    2021

  5. [5]

    D. M. Kennes, M. Claassen, L. Xian, A. Georges, A. J. Millis, J. Hone, C. R. Dean, D. N. Basov, A. N. Pasupa- thy, and A. Rubio, Nature Physics17, 155 (2021)

    2021

  6. [6]

    Claassen, L

    M. Claassen, L. Xian, D. M. Kennes, and A. Rubio, Na- ture Communications13, 4915 (2022)

    2022

  7. [7]

    F. Wu, L. Li, Q. Xu, L. Liu, Y. Yuan, J. Zhao, Z. Huang, X. Zan, K. Watanabe, T. Taniguchi, D. Shi, L. Xian, W. Yang, L. Du, and G. Zhang, Chin. Phys. Lett., 2023, Vol. 40, Issue 4, (2023)

    2023

  8. [8]

    C. S. Tsang, X. Zheng, T. Yang, Z. Yan, W. Han, L. W. Wong, H. Liu, S. Gao, K. H. Leung, C.-S. Lee, S. P. Lau, M. Yang, J. Zhao, and T. H. Ly, Science386, 198 (2024)

    2024

  9. [10]

    Bistritzer and A

    R. Bistritzer and A. H. MacDonald, Proceedings of the National Academy of Sciences108, 12233 (2011)

    2011

  10. [12]

    Pathak, T

    S. Pathak, T. Rakib, R. Hou, A. Nevidomskyy, E. Ertekin, H. T. Johnson, and L. K. Wagner, Physical Review B105, 10.1103/PhysRevB.105.115141 (2022)

    work page doi:10.1103/physrevb.105.115141 2022

  11. [13]

    C.-E. Ahn, W. Lee, K. Yananose, Y. Kim, and G. Y. Cho, Physical Review B110, 10.1103/PhysRevB.110.L161109 (2024)

    work page doi:10.1103/physrevb.110.l161109 2024

  12. [14]

    C. Xu, N. Mao, T. Zeng, and Y. Zhang, Physical Review Letters134, 10.1103/PhysRevLett.134.066601 (2025)

    work page doi:10.1103/physrevlett.134.066601 2025

  13. [15]

    Guinea and N

    F. Guinea and N. R. Walet, Physical Review B99, 10.1103/PhysRevB.99.205134 (2019)

    doi:10.1103/physrevb.99.205134 2019

  14. [16]

    K. T. Sch¨ utt, M. Gastegger, A. Tkatchenko, K.-R. M¨ uller, and R. J. Maurer, Nature Communications10, 5024 (2019)

    2019

  15. [17]

    O. T. Unke, M. Bogojeski, M. Gastegger, M. Geiger, T. Smidt, and K.-R. M¨ uller, SE(3)-equivariant predic- tion of molecular wavefunctions and electronic densities (2021), arXiv:2106.02347 [physics, stat]

    Pith/arXiv arXiv 2021

  16. [18]

    H. Li, Z. Wang, N. Zou, M. Ye, R. Xu, X. Gong, W. Duan, and Y. Xu, Nature Computational Science2, 367 (2022)

    2022

  17. [19]

    X. Gong, H. Li, N. Zou, R. Xu, W. Duan, and Y. Xu, Nature Communications14, 2848 (2023)

    2023

  18. [20]

    Zhong, H

    Y. Zhong, H. Yu, M. Su, X. Gong, and H. Xiang, npj Computational Materials9, 1 (2023)

    2023

  19. [21]

    Z. Tang, H. Li, P. Lin, X. Gong, G. Jin, L. He, H. Jiang, X. Ren, W. Duan, and Y. Xu, Nature Communications 15, 8815 (2024)

    2024

  20. [22]

    Y. Ma, H. Yu, Y. Zhong, S. Chen, X. Gong, and H. Xiang, Applied Physics Letters126, 044103 (2025)

    2025

  21. [23]

    Zhang, Y

    C. Zhang, Y. Zhong, Z.-G. Tao, X. Qin, H. Shang, Z. Lan, O. V. Prezhdo, X.-G. Gong, W. Chu, and H. Xiang, Na- ture Communications16, 2033 (2025)

    2033

  22. [24]

    Zhong, R

    Y. Zhong, R. Wang, X. Gong, and H. Xiang, A Universal Spin-Orbit-Coupled Hamiltonian Model for Accelerated Quantum Material Discovery (2025), arXiv:2504.19586 [cond-mat]

    Pith/arXiv arXiv 2025

  23. [25]

    A. M. Lewis, A. Grisafi, M. Ceriotti, and M. Rossi, Journal of Chemical Theory and Computation17, 7203 (2021)

    2021

  24. [26]

    P. B. Jørgensen and A. Bhowmik, npj Computational Materials8, 183 (2022)

    2022

  25. [28]

    Fiedler, N

    L. Fiedler, N. A. Modine, S. Schmerler, D. J. Vogel, G. A. Popoola, A. P. Thompson, S. Rajamanickam, and A. Cangi, npj Computational Materials9, 115 (2023)

    2023

  26. [29]

    T. Lv, Z. Zhong, Y. Liang, F. Li, J. Huang, and R. Zheng, Physical Review B108, 10.1103/PhysRevB.108.235159 (2023)

    work page doi:10.1103/physrevb.108.235159 2023

  27. [30]

    Rackers, L

    J. Rackers, L. Tecot, M. Geiger, and T. Smidt, Ma- chine Learning: Science and Technology 10.1088/2632- 2153/acb314 (2023)

    doi:10.1088/2632- 2023

  28. [31]

    Focassio, M

    B. Focassio, M. Domina, U. Patil, A. Fazzio, and S. San- vito, npj Computational Materials9, 87 (2023)

    2023

  29. [32]

    S. K. Achar, L. Bernasconi, and J. K. Johnson, Nanoma- terials13, 1853 (2023)

    2023

  30. [33]

    Lee and Y.-H

    R.-G. Lee and Y.-H. Kim, npj Computational Materials 10, 248 (2024)

    2024

  31. [34]

    Ushenin, K

    K. Ushenin, K. Khrabrov, A. Tsypin, A. Ber, E. Rumi- antsev, and A. Kadurin, Journal of Cheminformatics17, 65 (2025)

    2025

  32. [35]

    C. Li, O. Sharir, S. Yuan, and G. K.-L. Chan, Nature Communications16, 4811 (2025)

    2025

  33. [36]

    T. Bao, R. Xu, H. Li, X. Gong, Z. Tang, J. Fu, W. Duan, and Y. Xu, Deep-Learning Database of Density Functional Theory Hamiltonians for Twisted Materials (2024), arXiv:2404.06449 [cond-mat]

    Pith/arXiv arXiv 2024

  34. [37]

    Kohn, Physical Review Letters76, 3168 (1996)

    W. Kohn, Physical Review Letters76, 3168 (1996)

    1996

  35. [38]

    Behler, Chemical Reviews121, 10037 (2021)

    J. Behler, Chemical Reviews121, 10037 (2021)

    2021

  36. [39]

    Hohenberg, Physical Review136, B864 (1964)

    P. Hohenberg, Physical Review136, B864 (1964)

    1964

  37. [40]

    X. Ren, P. Rinke, V. Blum, J. Wieferink, A. Tkatchenko, A. Sanfilippo, K. Reuter, and M. Scheffler, New Journal of Physics14, 053020 (2012)

    2012

  38. [41]

    A. C. Ihrig, J. Wieferink, I. Y. Zhang, M. Ropo, X. Ren, P. Rinke, M. Scheffler, and V. Blum, New Journal of Physics17, 093020 (2015)

    2015

  39. [43]

    Grisafi, A

    A. Grisafi, A. Fabrizio, B. Meyer, D. M. Wilkins, C. Corminboeuf, and M. Ceriotti, ACS Central Science 5, 57 (2019)

    2019

  40. [44]

    A. M. Lewis, P. Lazzaroni, and M. Rossi, The Journal of Chemical Physics159, 014103 (2023)

    2023

  41. [45]

    Z. Y. Zhu, Y. C. Cheng, and U. Schwingenschl¨ ogl, Phys- ical Review B84, 10.1103/PhysRevB.84.153402 (2011)

    doi:10.1103/physrevb.84.153402 2011

  42. [47]

    Uchida, S

    K. Uchida, S. Furuya, J.-I. Iwata, and A. Oshiyama, Physical Review B90, 155451 (2014). 16

    2014

  43. [52]

    M. J. Calder´ on and E. Bascones, Physical Review B102, 155149 (2020)

    2020

  44. [53]

    Koshino, N

    M. Koshino, N. F. Q. Yuan, T. Koretsune, M. Ochi, K. Kuroki, and L. Fu, Physical Review X8, 031087 (2018)

    2018

  45. [54]

    Bennett, G

    D. Bennett, G. Chaudhary, R.-J. Slager, E. Bousquet, and P. Ghosez, Nature Communications14, 1629 (2023)

    2023

  46. [55]

    D. S. Kim, R. C. Dominguez, R. Mayorga-Luna, D. Ye, J. Embley, T. Tan, Y. Ni, Z. Liu, M. Ford, F. Y. Gao, S. Arash, K. Watanabe, T. Taniguchi, S. Kim, C.-K. Shih, K. Lai, W. Yao, L. Yang, X. Li, and Y. Miyahara, Nature Materials23, 65 (2024)

    2024

  47. [56]

    L. Gu, L. Zhang, S. Felsenfeld, B. Gao, R. Ma, S. Park, H. Jang, T. Taniguchi, K. Watanabe, and Y. Zhou, Phys- ical Review Letters135, 026901 (2025)

    2025

  48. [57]

    P. Zhao, C. Xiao, and W. Yao, npj 2D Materials and Applications5, 38 (2021)

    2021

  49. [58]

    Vachaspati,Kinks and Domain Walls: An Introduc- tion to Classical and Quantum Solitons(Cambridge Uni- versity Press, Cambridge, 2006)

    T. Vachaspati,Kinks and Domain Walls: An Introduc- tion to Classical and Quantum Solitons(Cambridge Uni- versity Press, Cambridge, 2006)

    2006

  50. [60]

    G. X. Ni, H. Wang, B.-Y. Jiang, L. X. Chen, Y. Du, Z. Y. Sun, M. D. Goldflam, A. J. Frenzel, X. M. Xie, M. M. Fogler, and D. N. Basov, Nature Communications 10, 4360 (2019)

    2019

  51. [61]

    C. R. Woods, P. Ares, H. Nevison-Andrews, M. J. Holwill, R. Fabregas, F. Guinea, A. K. Geim, K. S. Novoselov, N. R. Walet, and L. Fumagalli, Nature Com- munications12, 347 (2021)

    2021

  52. [62]

    Zhang, T

    Y. Zhang, T. Liu, and L. Fu, Physical Review B103, 10.1103/PhysRevB.103.155142 (2021)

    work page doi:10.1103/physrevb.103.155142 2021

  53. [63]

    Grisafi, J

    A. Grisafi, J. Nigam, and M. Ceriotti, Chemical Science 12, 2078 (2021)

    2078

  54. [64]

    J. Quan, N. Rybin, M. Scheffler, and C. Carbogno, Phys- ical Review B113, 085112 (2026)

    2026

  55. [65]

    C. Chen, K. P. Nuckolls, S. Ding, W. Miao, D. Wong, M. Oh, R. L. Lee, S. He, C. Peng, D. Pei, Y. Li, C. Hao, H. Yan, H. Xiao, H. Gao, Q. Li, S. Zhang, J. Liu, L. He, K. Watanabe, T. Taniguchi, C. Jozwiak, A. Bostwick, E. Rotenberg, C. Li, X. Han, D. Pan, Z. Liu, X. Dai, C. Liu, B. A. Bernevig, Y. Wang, A. Yazdani, and Y. Chen, Nature636, 342 (2024)

    2024

  56. [66]

    C. Chen, W. Holtzmann, X.-W. Zhang, E. Anderson, S. He, Y. Zhao, W. Li, J. Liu, Y. Guo, C. Jozwiak, A. Bostwick, E. Rotenberg, K. Watanabe, T. Taniguchi, T. Cao, D. Xiao, X. Xu, and Y. Chen, Communications Physics 10.1038/s42005-026-02497-8 (2026)

    work page doi:10.1038/s42005-026-02497-8 2026

  57. [67]

    A. P. Bart´ ok, R. Kondor, and G. Cs´ anyi, Physical Review B87, 184115 (2013)

    2013

  58. [68]

    M. J. Willatt, F. Musil, and M. Ceriotti, The Journal of Chemical Physics150, 154110 (2019)

    2019

  59. [70]

    G. H. Golub and C. F. Van Loan,Matrix Computations, fourth edition ed., Johns Hopkins Studies in the Math- ematical Sciences (The Johns Hopkins University Press, Baltimore, 2013)

    2013

  60. [71]

    M. P. Deisenroth, A. A. Faisal, and C. S. Ong,Mathemat- ics for Machine Learning(Cambridge University Press, Cambridge, UK New York, NY, 2020)

    2020

  61. [72]

    V. Blum, R. Gehrke, F. Hanke, P. Havu, V. Havu, X. Ren, K. Reuter, and M. Scheffler, Computer Physics Communications180, 2175 (2009)

    2009

  62. [75]

    Stacking Symmetry

    J. P. Perdew, K. Burke, and M. Ernzerhof, Physical Re- view Letters77, 3865 (1996). 18 Long-Range Machine Learning of Electron Density for Twisted Bilayer Moir´ e Materials SUPPORTING INFORMATION Zekun Lou,1 Alan M. Lewis, 2 and Mariana Rossi 1,3 1MPI for the Structure and Dynamics of Matter, Luruper Chaussee 149, 22761 Hamburg, Germany 2Department of Che...

    1996

  63. [76]

    V. Blum, R. Gehrke, F. Hanke, P. Havu, V. Havu, X. Ren, K. Reuter, and M. Scheffler, Computer Physics Communications 180, 2175 (2009)

    2009

  64. [77]

    J. W. Abbott, C. M. Acosta, A. Akkoush, A. Ambrosetti, V. Atalla, A. Bagrets, J. Behler, D. Berger, B. Bieniek, J. Bj¨ ork, V. Blum, S. Bohloul, C. L. Box, N. Boyer, D. S. Brambila, G. A. Bramley, K. R. Bryenton, M. Camarasa-G´ omez, C. Carbogno, F. Caruso, S. Chutia, M. Ceriotti, G. Cs´ anyi, W. Dawson, F. A. Delesma, F. D. Sala, B. Delley, R. A. D. Jr, ...

    Pith/arXiv arXiv 2025

  65. [78]

    K. V. Zakharchenko, M. I. Katsnelson, and A. Fasolino, Physical Review Letters102, 10.1103/PhysRevLett.102.046808 (2009)

    work page doi:10.1103/physrevlett.102.046808 2009

  66. [79]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, APL Materials1, 011002 (2013)

    2013

  67. [80]

    S. M. Gilbert, T. Pham, M. Dogan, S. Oh, B. Shevitski, G. Schumm, S. Liu, P. Ercius, S. Aloni, M. L. Cohen, and A. Zettl, 2D Materials6, 021006 (2019)

    2019

  68. [81]

    J. T. Krogel, S. F. Yuk, P. R. C. Kent, and V. R. Cooper, The Journal of Physical Chemistry A124, 9867 (2020)

    2020

  69. [82]

    Kumar, S

    Anisha, R. Kumar, S. Srivastava, and T. Kumar, Physica Scripta99, 015914 (2023)

    2023

  70. [83]

    J. He, K. Hummer, and C. Franchini, Physical Review B89, 075409 (2014)

    2014

  71. [84]

    Tkatchenko, Physical Review Letters108, 10.1103/PhysRevLett.108.236402 (2012)

    A. Tkatchenko, Physical Review Letters108, 10.1103/PhysRevLett.108.236402 (2012)

    doi:10.1103/physrevlett.108.236402 2012

  72. [85]

    Hermann and A

    J. Hermann and A. Tkatchenko, Physical Review Letters124, 146401 (2020)

    2020

  73. [86]

    J. M. B. Lopes dos Santos, N. M. R. Peres, and A. H. Castro Neto, Physical Review B86, 155449 (2012)

    2012

  74. [87]

    J. P. Perdew, K. Burke, and M. Ernzerhof, Physical Review Letters77, 3865 (1996)

    1996

  75. [88]

    Grisafi, D

    A. Grisafi, D. M. Wilkins, G. Cs´ anyi, and M. Ceriotti, Physical Review Letters120, 036002 (2018)

    2018

  76. [89]

    Grisafi, A

    A. Grisafi, A. M. Lewis, M. Rossi, and M. Ceriotti, Journal of Chemical Theory and Computation19, 4451 (2023)

    2023

  77. [90]

    Grisafi and M

    A. Grisafi and M. Ceriotti, The Journal of Chemical Physics151, 204105 (2019)

    2019

  78. [91]

    Tanwar, U

    P. Tanwar, U. Paliwal, K. B. Joshi, and J. Kumar, Journal of Physics and Chemistry of Solids179, 111382 (2023)

    2023

  79. [92]

    W. P. Huhn and V. Blum, Physical Review Materials1, 10.1103/PhysRevMaterials.1.033803 (2017)

    doi:10.1103/physrevmaterials.1.033803 2017

  80. [93]

    Batatia, D

    I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, and G. Csanyi, Advances in Neural Information Processing Systems35, 11423 (2022)

    2022

Showing first 80 references.