Pith. sign in

REVIEW 2 major objections 4 minor 3 cited by

A neural-network variational wavefunction, trained purely by energy minimization, discovers a fractional Chern insulator in a zero-net-flux periodic magnetic field, and a new 'momentum spectroscopy' protocol extracts its threefold topologic

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 19:06 UTC pith:PQ623AHG

load-bearing objection NN-VMC discovers an FCI in a zero-net-flux continuum model and momentum spectroscopy reads out a threefold degeneracy, but the diagnostic's specificity is unproven without a control; still worth a serious referee. the 2 major comments →

arxiv 2512.01863 v2 pith:PQ623AHG submitted 2025-12-01 cond-mat.mes-hall cond-mat.str-elcs.AI

Topological Order in Neural Wavefunctions

Ahmed Abouelkomsan , Max Geier , Liang Fu This is my paper
classification cond-mat.mes-hall cond-mat.str-elcs.AI
keywords topological orderfractional Chern insulatorneural network variational Monte Carlomomentum spectroscopyself-attentioncharge density wavezero net fluxflat Chern band
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 claims that a self-attention-based neural network can discover a topologically ordered fractional Chern insulator ground state using nothing but energy minimization: no band structure, no Chern number, and no symmetry information is fed in. The key technical advance is 'momentum spectroscopy,' a cheap post-processing step that projects the single optimized wavefunction onto center-of-mass momentum sectors, yielding the three quasi-degenerate ground states whose existence is the hallmark of topological order. Applied to a continuum model of spinless fermions in a periodic magnetic field with zero net flux — a regime where fractionalization was previously unclear — the method finds a clean gapped liquid at ν=1/3 with threefold degeneracy, and it also finds the competing charge-density-wave when the modulation is stronger. If correct, this establishes neural-network variational Monte Carlo as a practical tool for discovering strongly correlated topological phases without any prior bias.

Core claim

The authors demonstrate that an attention-based neural quantum state, optimized purely to minimize the variational energy of a continuum model of spinless fermions in a periodic magnetic field with zero net flux, converges to a featureless gapped quantum liquid at filling ν=1/3. The optimized wavefunction has nonzero weight in exactly three center-of-mass momentum sectors, and the variational energies in those sectors are quasi-degenerate; these three momenta coincide with the known FCI ground-state momenta obtained from generalized Pauli-principle counting rules. This, together with the saturated structure-factor bound and the absence of Bragg peaks, identifies the state as a fractional Che

What carries the argument

The central object is the momentum-spectroscopy protocol: the optimized real-space wavefunction Ψ({rᵢ}) is expanded in eigenstates Φ_K of the center-of-mass translation operator T(R) via a Fourier projection Φ_K = (1/N_s) Σ_R e^{-iK·R} Ψ({rᵢ+R}), and the weights |c_K|² together with the projected energies E_K = ⟨Φ_K|H|Φ_K⟩ are computed. Because the neural network is translationally invariant in its parametrization, a state that has converged to the true ground-state manifold must have weight only in the momentum sectors K_top of the degenerate ground states, and those sectors' energies must be quasi-degenerate. This converts a single momentum-agnostic optimization into a full spectroscopic d

Load-bearing premise

The result hinges on the assumption that a converged neural wavefunction has zero weight in all momentum sectors except the degenerate ground-state sectors and that the projected energies there are quasi-degenerate; no control experiment is shown.

What would settle it

A concrete test: run the same momentum spectroscopy on a model known to be a trivial insulator or on a symmetry-broken CDW; if the optimized wavefunction also shows non-negligible weights in multiple momentum sectors with quasi-degenerate energies, the diagnostic would not be specific to topological order. Alternatively, check whether the three projected states at ν=1/3 remain exactly degenerate and distinct as system size and torus aspect ratio are varied, and whether the Hall conductance (via flux insertion) is quantized.

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

If this is right

  • Neural-network variational Monte Carlo can discover fractional Chern insulators from scratch in continuum models, without band projection or any topological input, meaning it can be applied to realistic moiré models where multiband effects matter.
  • Momentum spectroscopy detects topological ground-state degeneracy from a single optimized wavefunction at negligible extra cost, avoiding separate optimizations in each momentum sector.
  • The same ansatz captures both a fractional Chern insulator and a competing charge-density wave at different parameter values, establishing that the method is unbiased enough to map out competing orders.
  • Because the calculation is performed in real space, it naturally includes all energy bands and finds variational energies below band-projected exact diagonalization, opening the door to larger system sizes.

Where Pith is reading between the lines

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

  • A natural extension is to apply momentum spectroscopy as a diagnostic to any variational wavefunction, not just neural ones; the protocol's assumption that converged states have zero weight in other sectors could be tested on a trivial insulator to establish a control.
  • If the method generalizes, it suggests that topological order can be discovered by energy minimization alone in models with zero net flux, which would significantly broaden the search space for fractionalized phases in moiré and strained materials.
  • The approach could be extended to detect non-Abelian topological order by looking for degeneracies equal to the number of anyon types, though the protocol requires that degenerate states carry distinct momenta or other symmetries.
  • A testable prediction: performing the same momentum projection on a symmetry-broken CDW should yield multiple sectors as well, so the degeneracy alone is not sufficient; the combination of liquid density, structure factor, and quasi-degenerate projected energies is what pins down topological order.

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

2 major / 4 minor

Summary. The paper develops a momentum-spectroscopy protocol for detecting ground-state topological degeneracy from a single optimized neural-network variational wavefunction. It applies this to a continuum model of spinless fermions in a zero-net-flux periodic magnetic field. At filling ν=1/3 and λ=-0.23, the optimized attention-based NN wavefunction appears translationally invariant and liquid-like; after projecting onto center-of-mass momentum sectors, the wavefunction has nonzero weight in only three momentum sectors, with quasi-degenerate projected variational energies. The authors claim this threefold degeneracy is the topological ground-state degeneracy of a fractional Chern insulator, discovered without prior band-structure information. At λ=-0.26 the same ansatz yields a charge density wave. The paper also compares NN energies to band-projected exact diagonalization and supports the FCI interpretation with the structure-factor quantum-weight bound.

Significance. If the central claim holds, this is a valuable advance: it demonstrates that a general-purpose NN-VMC ansatz can discover a fractional Chern insulator in a continuum model without band projection, and it introduces a post-processing diagnostic that extracts topological degeneracy from a single variational state. The model studied is interesting and the paper is clearly written, with detailed architecture, hyperparameters, training curves, and ED comparisons in the supplementary material. The momentum-spectroscopy idea is potentially widely applicable. However, two load-bearing pillars — the specificity of the diagnostic and the accuracy benchmark for the variational state — are not yet established, and both are directly testable.

major comments (2)
  1. [Momentum spectroscopy (Eqs. (1)–(2), Fig. 4)] The protocol's central assumption — that an accurate variational ground state has nonzero weights only in the degenerate ground-state momentum sectors — is stated but not demonstrated. The Psiformer ansatz is optimized without translation symmetry, so small spurious weights in other sectors are inevitable. No threshold, statistical error, or convergence criterion is given for Fig. 4(a). The Discussion explicitly concedes the method is 'not unique to topologically ordered states' and can also detect generic low-lying excitations. A CDW also has degenerate ground states at different momenta, and a trivial insulator could in principle produce small weights in several sectors. Please add control experiments: apply the same protocol to the λ=-0.26 CDW state and to a known trivial state, and show that the uniform density plus exactly three quasi-degenerate sectors is specific to the FCI. Quant
  2. [Results and Supplementary Sec. D (Table II, Fig. S2)] The claim that NN-VMC 'achieves lower energy than ED projected onto the lowest band' is not a meaningful accuracy benchmark, because the NN ansatz is not restricted to the lowest-band Hilbert space; lower energy is expected by construction. Since the momentum-spectroscopy inference relies on Ψ being an accurate approximation to the true ground state, the paper should provide a more direct accuracy check. For the smallest system (N=3 in 9 unit cells), a full unprojected ED comparison should be feasible, or the authors should present other evidence (e.g., systematic convergence of the energy and momentum weights). Without this, the 'remarkable accuracy' statement is not quantitatively supported.
minor comments (4)
  1. [Results (Fig. 3(d))] The notation '|q|^2 A/4π S(q)' is ambiguous. The intended quantity appears to be |q|^2 A / (4π S(q)), which approaches 3 at small |q|. Please write the formula explicitly.
  2. [Throughout] Several typos: 'controlled by the dielectric constant' should be 'controlled'; 'umambigously' should be 'unambiguously'; the main-text title 'Topological Order in Deep State' differs from the arXiv title 'Topological Order in Neural Wavefunctions'.
  3. [References] Reference [63] (Adam) has a corrupted arXiv identifier: 'arXiv:1412.69801412' should be 'arXiv:1412.6980'.
  4. [Supplementary Fig. S2 / Table II] The ED comparison is shown for N=3 in 9 cells, while the main results use N=8 in 24 cells and N=9 in 27 cells. Clarify why the larger systems are beyond multiband ED and whether the smaller-system comparison is representative of the physics at the sizes used for the topological-degeneracy diagnostic.

Circularity Check

0 steps flagged

No significant circularity: the FCI discovery is energy-minimized and benchmarked; momentum spectroscopy is a post-hoc readout, not a fit, and cited inputs are external theorems or independent comparisons.

full rationale

The central derivation chain—NN-VMC energy minimization of Hamiltonian (3), followed by momentum projection of the optimized wavefunction and quasi-degenerate energies—does not reduce to its inputs by construction. The weights |<Psi|Phi_K>|^2 are computed from the optimized state rather than fitted to force three sectors; the three-sector result is an empirical property of the converged Psi. The identification of those sectors as FCI ground-state momenta uses external counting rules (refs 9,45-47), and the structure-factor check uses an external theorem (refs 42,43); neither is an input to the variational ansatz, and the paper benchmarks energies against independent ED. The paper's own Discussion concedes the diagnostic is not unique to topological order ('this approach is not unique to topologically ordered states but can be used to detect generic low-lying excitations above the ground states'), and no control experiment is shown; this is a specificity/validation caveat, not a circularity. Self-citations to Psiformer and the topological bound are load-bearing but are supported by external benchmarking and parameter-free theorems, so they do not raise the circularity score.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The model parameters λ and r_s are physical inputs of the model, not numbers fitted to produce the claimed degeneracy. The NN parameters are optimized against the Hamiltonian, which is the VMC method itself; no additional fit parameters are introduced to force topological order. The load-bearing premises are the diagnostic assumption and the expressivity/symmetry assumptions above.

axioms (5)
  • domain assumption The continuum Hamiltonian (3) with periodic C6-symmetric magnetic field (4) and Coulomb interaction is a valid model for the zero-net-flux periodic-field system; its lowest band has Chern number C=1 and is flat for λ=-0.23.
    The model is motivated by the adiabatic limit of twisted TMDs (Refs 31-34); the band structure is presented in Fig 2. The central simulation uses this Hamiltonian as the ground-truth system.
  • domain assumption The Psiformer self-attention ansatz (Eq. 5 and SM Eq. A13) is expressive enough to represent the FCI ground state accurately after energy minimization.
    Standard expressivity assumption for neural quantum states; no proof of universal approximation for this architecture is given.
  • ad hoc to paper Momentum spectroscopy: an accurate variational ground state has nonzero COM-momentum weights only in the degenerate ground-state sectors, and the projected energies are quasi-degenerate.
    This is the core diagnostic premise, asserted rather than derived or benchmarked. The paper's discovery of topological order rests on this premise.
  • standard math The structure-factor topological bound K ≥ A|C|/(4π) (Refs 42,43) applies to the unprojected structure factor computed from the NN wavefunction.
    Used in Fig 3(d) to argue consistency with a gapped topological phase.
  • standard math The expected FCI ground-state momenta for this finite system are given by generalized Pauli principle / thin-torus counting rules (Refs 9,45-47).
    Used to identify the three momentum sectors found in Fig 4(a) as the FCI sectors.

pith-pipeline@v1.3.0-alltime-deepseek · 13322 in / 18454 out tokens · 184605 ms · 2026-08-03T19:06:08.328959+00:00 · methodology

0 comments
read the original abstract

Topologically ordered states are among the most interesting quantum phases of matter that host emergent quasi-particles having fractional charge and obeying fractional quantum statistics. Theoretical study of such states is however challenging owing to their strong-coupling nature that prevents conventional mean-field treatment. Here, we demonstrate that an attention-based deep neural network provides an expressive variational wavefunction that discovers fractional Chern insulator ground states purely through energy minimization without prior knowledge and achieves remarkable accuracy. We introduce an efficient method to extract ground state topological degeneracy -- a hallmark of topological order -- from a single optimized real-space wavefunction in translation-invariant systems by decomposing it into different many-body momentum sectors. Our results establish neural network variational Monte Carlo as a versatile tool for discovering strongly correlated topological phases.

Figures

Figures reproduced from arXiv: 2512.01863 by Ahmed Abouelkomsan, Liang Fu, Max Geier.

Figure 2
Figure 2. Figure 2: FIG. 2. Band structure of the model defined in equation (3) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: FIG. 1. (a) Schematic of the self-attention–based neural [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. (a) The overlap [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Ground state charge density [PITH_FULL_IMAGE:figures/full_fig_p004_3.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 3 Pith papers

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

  1. Fermi Sets: Universal and interpretable neural architectures for fermions

    cond-mat.str-el 2026-01 unverdicted novelty 7.0

    Fermi Sets achieve universal approximation of fermionic wavefunctions using K antisymmetric bases times symmetric neural networks, where K equals 1 in 1D, 2 in 2D, and grows linearly with particle number in higher dimensions.

  2. Topological invariant of periodic many body wavefunction from charge pumping simulation

    cond-mat.str-el 2026-04 unverdicted novelty 6.0

    Charge-pumping simulation extracts Chern numbers and identifies anomalous composite Fermi liquids from neural network wavefunctions in fractional Chern insulators.

  3. Large Electron Model: A Universal Ground State Predictor

    cond-mat.str-el 2026-03 conditional novelty 6.0

    A parameter-conditioned Fermi Sets neural network, trained by variational Monte Carlo on a small grid of (λ,N), predicts ground-state wavefunctions of 2D quantum dots for unseen interaction strengths and particle numb...

Reference graph

Works this paper leans on

68 extracted references · 4 linked inside Pith · cited by 3 Pith papers

  1. [1]

    H. Park, J. Cai, E. Anderson, Y. Zhang, J. Zhu, X. Liu, C. Wang, W. Holtzmann, C. Hu, Z. Liu, et al., Observa- tion of fractionally quantized anomalous hall effect, Na- ture622, 74 (2023)

    2023

  2. [2]

    J. Cai, E. Anderson, C. Wang, X. Zhang, X. Liu, W. Holtzmann, Y. Zhang, F. Fan, T. Taniguchi, K. Watanabe, et al., Signatures of fractional quantum anomalous hall states in twisted mote2, Nature622, 63 (2023)

    2023

  3. [3]

    Y. Zeng, Z. Xia, K. Kang, J. Zhu, P. Kn¨ uppel, C. Vaswani, K. Watanabe, T. Taniguchi, K. F. Mak, and J. Shan, Thermodynamic evidence of fractional chern in- sulator in moir´ e mote2, Nature622, 69 (2023)

    2023

  4. [4]

    F. Xu, Z. Sun, T. Jia, C. Liu, C. Xu, C. Li, Y. Gu, K. Watanabe, T. Taniguchi, B. Tong, et al., Observation of integer and fractional quantum anomalous hall effects in twisted bilayer mote 2, Physical Review X13, 031037 (2023)

    2023

  5. [5]

    Z. Lu, T. Han, Y. Yao, A. P. Reddy, J. Yang, J. Seo, K. Watanabe, T. Taniguchi, L. Fu, and L. Ju, Fractional quantum anomalous hall effect in multilayer graphene, Nature626, 759 (2024)

    2024

  6. [6]

    Wen, Quantum field theory of many-body systems: From the origin of sound to an origin of light and elec- trons, (2004)

    X.-G. Wen, Quantum field theory of many-body systems: From the origin of sound to an origin of light and elec- trons, (2004)

    2004

  7. [7]

    Tang, J.-W

    E. Tang, J.-W. Mei, and X.-G. Wen, High-temperature fractional quantum hall states, Physical review letters 106, 236802 (2011)

    2011

  8. [8]

    Neupert, L

    T. Neupert, L. Santos, C. Chamon, and C. Mudry, Frac- tional quantum hall states at zero magnetic field, Physi- cal review letters106, 236804 (2011)

    2011

  9. [9]

    Regnault and B

    N. Regnault and B. A. Bernevig, Fractional chern insu- lator, Physical Review X1, 021014 (2011)

    2011

  10. [10]

    Sheng, Z.-C

    D. Sheng, Z.-C. Gu, K. Sun, and L. Sheng, Fractional quantum hall effect in the absence of landau levels, Na- ture communications2, 389 (2011)

    2011

  11. [11]

    Abouelkomsan, Z

    A. Abouelkomsan, Z. Liu, and E. J. Bergholtz, Particle- hole duality, emergent fermi liquids, and fractional chern insulators in moir´ e flatbands, Physical review letters124, 106803 (2020)

    2020

  12. [12]

    Repellin and T

    C. Repellin and T. Senthil, Chern bands of twisted bi- layer graphene: Fractional chern insulators and spin 6 phase transition, Physical Review Research2, 023238 (2020)

    2020

  13. [13]

    P. J. Ledwith, G. Tarnopolsky, E. Khalaf, and A. Vish- wanath, Fractional Chern insulator states in twisted bi- layer graphene: An analytical approach, Physical Review Research2, 023237 (2020)

    2020

  14. [14]

    H. Li, U. Kumar, K. Sun, and S.-Z. Lin, Spontaneous fractional chern insulators in transition metal dichalco- genide moir´ e superlattices, Physical Review Research3, L032070 (2021)

    2021

  15. [15]

    Cr´ epel and L

    V. Cr´ epel and L. Fu, Anomalous hall metal and fractional chern insulator in twisted transition metal dichalco- genides, Physical Review B107, L201109 (2023)

    2023

  16. [16]

    Abouelkomsan, A

    A. Abouelkomsan, A. P. Reddy, L. Fu, and E. J. Bergholtz, Band mixing in the quantum anomalous hall regime of twisted semiconductor bilayers, Physical Re- view B109, L121107 (2024)

    2024

  17. [17]

    J. Yu, J. Herzog-Arbeitman, M. Wang, O. Vafek, B. A. Bernevig, and N. Regnault, Fractional chern insulators versus nonmagnetic states in twisted bilayer mote 2, Physical Review B109, 045147 (2024)

    2024

  18. [18]

    Carleo and M

    G. Carleo and M. Troyer, Solving the quantum many- body problem with artificial neural networks, Science 355, 602 (2017)

    2017

  19. [19]

    W. M. Foulkes, L. Mitas, R. Needs, and G. Rajagopal, Quantum monte carlo simulations of solids, Reviews of Modern Physics73, 33 (2001)

    2001

  20. [20]

    Y. Teng, D. D. Dai, and L. Fu, Solving the fractional quantum hall problem with self-attention neural network, Physical Review B111, 205117 (2025)

    2025

  21. [21]

    C.-T. Li, T. Ong, M. Geier, H. Lin, and L. Fu, Attention is all you need to solve chiral superconductivity, arXiv preprint arXiv:2509.03683 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    Kitaev and J

    A. Kitaev and J. Preskill, Topological entanglement en- tropy, Physical review letters96, 110404 (2006)

    2006

  23. [23]

    Levin and X.-G

    M. Levin and X.-G. Wen, Detecting topological order in a ground state wave function, Physical review letters96, 110405 (2006)

    2006

  24. [24]

    M. B. Hastings, I. Gonz´ alez, A. B. Kallin, and R. G. Melko, Measuring renyi entanglement entropy in quan- tum monte carlo simulations, Physical review letters104, 157201 (2010)

    2010

  25. [25]

    Hibat-Allah, M

    M. Hibat-Allah, M. Ganahl, L. E. Hayward, R. G. Melko, and J. Carrasquilla, Recurrent neural network wave func- tions, Physical Review Research2, 023358 (2020)

    2020

  26. [26]

    Hibat-Allah, R

    M. Hibat-Allah, R. G. Melko, and J. Carrasquilla, In- vestigating topological order using recurrent neural net- works, Physical Review B108, 075152 (2023)

    2023

  27. [27]

    Wang and E

    Z. Wang and E. J. Davis, Calculating r´ enyi entropies with neural autoregressive quantum states, Physical Review A 102, 062413 (2020)

    2020

  28. [28]

    Wen and Q

    X.-G. Wen and Q. Niu, Ground-state degeneracy of the fractional quantum hall states in the presence of a ran- dom potential and on high-genus riemann surfaces, Phys- ical Review B41, 9377 (1990)

    1990

  29. [29]

    Q. Niu, D. J. Thouless, and Y.-S. Wu, Quantized hall conductance as a topological invariant, Physical Review B31, 3372 (1985)

    1985

  30. [30]

    See Supplementary Material for additional details about the NN architecture including benchmarking with ED

  31. [31]

    F. Wu, T. Lovorn, E. Tutuc, I. Martin, and A. MacDon- ald, Topological insulators in twisted transition metal dichalcogenide homobilayers, Physical review letters122, 086402 (2019)

    2019

  32. [32]

    N. Paul, Y. Zhang, and L. Fu, Giant proximity exchange and flat chern band in 2d magnet-semiconductor het- erostructures, Science Advances9, eabn1401 (2023)

    2023

  33. [33]

    Morales-Dur´ an, N

    N. Morales-Dur´ an, N. Wei, J. Shi, and A. H. MacDonald, Magic angles and fractional chern insulators in twisted homobilayer transition metal dichalcogenides, Physical Review Letters132, 096602 (2024)

    2024

  34. [34]

    J. Shi, N. Morales-Dur´ an, E. Khalaf, and A. H. MacDon- ald, Adiabatic approximation and aharonov-casher bands in twisted homobilayer transition metal dichalcogenides, Physical Review B110, 035130 (2024)

    2024

  35. [35]

    Li and F

    B. Li and F. Wu, Variational mapping of chern bands to landau levels: Application to fractional chern insulators in twisted mote 2, Physical Review B111, 125122 (2025)

    2025

  36. [36]

    A. P. Reddy, N. Paul, A. Abouelkomsan, and L. Fu, Non- abelian fractionalization in topological minibands, Phys- ical review letters133, 166503 (2024)

    2024

  37. [37]

    A. P. Reddy, D. Sheng, A. Abouelkomsan, E. J. Bergholtz, and L. Fu, Anti-topological crystal and non- abelian liquid in twisted semiconductor bilayers, arXiv preprint arXiv:2411.19898 (2024)

    Pith/arXiv arXiv 2024

  38. [38]

    D. Pfau, J. S. Spencer, A. G. Matthews, and W. M. C. Foulkes, Ab initio solution of the many-electron schr¨ odinger equation with deep neural networks, Physical review research2, 033429 (2020)

    2020

  39. [39]

    von Glehn, J

    I. von Glehn, J. S. Spencer, and D. Pfau, A self-attention ansatz for ab-initio quantum chemistry, arXiv preprint arXiv:2211.13672 (2022)

    Pith/arXiv arXiv 2022

  40. [40]

    Geier, K

    M. Geier, K. Nazaryan, T. Zaklama, and L. Fu, Self- attention neural network for solving correlated elec- tron problems in solids, Physical Review B112, 045119 (2025)

    2025

  41. [41]

    Nazaryan, F

    K. Nazaryan, F. Gaggioli, Y. Teng, and L. Fu, Artificial intelligence for quantum matter: Finding a needle in a haystack, arXiv preprint arXiv:2507.13322 (2025)

    arXiv 2025

  42. [42]

    Onishi and L

    Y. Onishi and L. Fu, Topological bound on the structure factor, Physical Review Letters133, 206602 (2024)

    2024

  43. [43]

    Zaklama, D

    T. Zaklama, D. Luo, and L. Fu, Structure factor and topological bound of twisted bilayer semiconductors at fractional fillings, Physical Review B112, L041115 (2025)

    2025

  44. [44]

    R. P. Feynman, Atomic theory of the two-fluid model of liquid helium, Physical Review94, 262 (1954)

    1954

  45. [45]

    Bergholtz and A

    E. Bergholtz and A. Karlhede, Quantum hall system in tao-thouless limit, Physical Review B—Condensed Mat- ter and Materials Physics77, 155308 (2008)

    2008

  46. [46]

    B. A. Bernevig and F. Haldane, Model fractional quan- tum hall states and jack polynomials, Physical review letters100, 246802 (2008)

    2008

  47. [47]

    B. A. Bernevig and N. Regnault, Emergent many-body translational symmetries of abelian and non-abelian frac- tionally filled topological insulators, Physical Review B—Condensed Matter and Materials Physics85, 075128 (2012)

    2012

  48. [48]

    X. Li, Y. Chen, B. Li, H. Chen, F. Wu, J. Chen, and W. Ren, Deep learning sheds light on integer and fractional topological insulators, arXiv preprint arXiv:2503.11756 (2025)

    Pith/arXiv arXiv 2025

  49. [49]

    Zhang and D

    L. Zhang and D. Luo, Neural transformer backflow for solving momentum-resolved ground states of strongly correlated materials, arXiv preprint arXiv:2509.09275 (2025)

    arXiv 2025

  50. [50]

    Kobayashi, Y

    R. Kobayashi, Y. Zhang, N. Manjunath, and M. Barkeshli, Crystalline invariants of fractional 7 chern insulators, Physical Review B112, 035147 (2025)

    2025

  51. [51]

    K. Choo, G. Carleo, N. Regnault, and T. Neupert, Sym- metries and many-body excitations with neural-network quantum states, Physical review letters121, 167204 (2018)

    2018

  52. [52]

    Tang and L

    E. Tang and L. Fu, Strain-induced partially flat band, helical snake states and interface superconductivity in topological crystalline insulators, Nature Physics10, 964 (2014)

    2014

  53. [53]

    Tahir, O

    M. Tahir, O. Pinaud, and H. Chen, Emergent flat band lattices in spatially periodic magnetic fields, Physical Re- view B102, 035425 (2020)

    2020

  54. [54]

    Q. Gao, J. Dong, P. Ledwith, D. Parker, and E. Khalaf, Untwisting moir´ e physics: Almost ideal bands and frac- tional chern insulators in periodically strained monolayer graphene, Physical Review Letters131, 096401 (2023)

    2023

  55. [55]

    Aidelsburger, M

    M. Aidelsburger, M. Atala, M. Lohse, J. T. Barreiro, B. Paredes, and I. Bloch, Realization of the hofstadter hamiltonian with ultracold atoms in optical lattices, Physical review letters111, 185301 (2013)

    2013

  56. [56]

    Abouelkomsan and L

    A. Abouelkomsan and L. Fu, Non-abelian spin hall insu- lator, Physical Review Research7, 023083 (2025)

    2025

  57. [57]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, At- tention is all you need, Advances in neural information processing systems30(2017)

    2017

  58. [58]

    Becca and S

    F. Becca and S. Sorella, Quantum Monte Carlo approaches for correlated systems (Cambridge University Press, 2017)

    2017

  59. [59]

    Sorella, Green function monte carlo with stochastic reconfiguration, Physical review letters80, 4558 (1998)

    S. Sorella, Green function monte carlo with stochastic reconfiguration, Physical review letters80, 4558 (1998)

    1998

  60. [60]

    Amari, Natural gradient works efficiently in learn- ing, Neural computation10, 251 (1998)

    S.-I. Amari, Natural gradient works efficiently in learn- ing, Neural computation10, 251 (1998)

    1998

  61. [61]

    Stokes, J

    J. Stokes, J. Izaac, N. Killoran, and G. Carleo, Quantum natural gradient, Quantum4, 269 (2020)

    2020

  62. [62]

    Martens and R

    J. Martens and R. Grosse, Optimizing neural net- works with kronecker-factored approximate curvature, in International conference on machine learning (PMLR,

  63. [63]

    Topological Order in Deep State

    K. D. B. J. Adam et al., A method for stochastic opti- mization, arXiv preprint arXiv:1412.69801412(2014). 1 Supplementary material for: “Topological Order in Deep State” Ahmed Abouelkomsan1, Max Geier 1 and Liang Fu1 1Department of Physics, Massachusetts Institute of Technology, Cambridge, MA-02139, USA In the supplementary material, we provide details a...

    arXiv 2014

  64. [64]

    Multilayer Perceptron (MLP) MLPs are standard feed-forward neural networks that implements the following transformation on agenericvector g1 ∈R dL to yield another vectorg 2, g2 =g 1 +F(Wg 1 +b) (A3) withWis a linear transformationW∈R dL ×R dL andb∈R dL is a bias vector.Fhere represents a non-linear activation function which we choose to be the GELU function

  65. [65]

    2 in the main text), MLPs act individually on each particle without mixing different particle streams which cannot describe correlations between particles

    Self-attention In the NN architecture (Fig. 2 in the main text), MLPs act individually on each particle without mixing different particle streams which cannot describe correlations between particles. In order to capture such correlations, we utilize the self-attention mechanism [57] which form the basis of transformers used in large language models. Self-...

  66. [66]

    ϕn j (ri;{r ̸=i}) =w n 2j ·h L i +iw n 2j+1 ·h L i (A12) wherew n 2j andw n 2j+1 are projection matrices that construct the real and the imaginary part of the generalized orbitals

    Projection ontoN det The final output of the neural network{hL i }is projected onton= 1,· · ·, Ndet distinct sets of generalized many-body orbitals. ϕn j (ri;{r ̸=i}) =w n 2j ·h L i +iw n 2j+1 ·h L i (A12) wherew n 2j andw n 2j+1 are projection matrices that construct the real and the imaginary part of the generalized orbitals. The full wavefunction ansat...

  67. [67]

    All calculations were done for rs ≈3.43

    computed from band projected ED and NN-VMC. All calculations were done for rs ≈3.43. three particles in 9 unit cells. Increasing the number of bandsN b, the ED energies approaches the true energy value of the systems. However, the NN energy is still very close showing that it is an excellent approximation for the ground state of the system. In Table. II, ...

  68. [68]

    Energy at a certain step is averaged over the previous 4000 steps

    as a function of training steps forN= 8 ,N s = 24 andλ=−0.23. Energy at a certain step is averaged over the previous 4000 steps. The green line denotes 1 band projected ED. The vertical dashed lines denote instances when the training was stopped and resumed with smaller learning rates. In Fig. S3, we show a training curve obtained forN= 8 andλ=−0.23 when ...