Pith. sign in

REVIEW 3 major objections 6 minor 48 references

Explorative Curriculum Learning for Strongly Correlated Electron Systems

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A curriculum that fine-tunes neural quantum states step by step to the nearest parameter value speeds up Hubbard-model exploration by over 200-fold.

desk verdict Nearest-neighbor curriculum ordering for NQS transfer is a genuinely useful idea with credible benchmark support, but the headline 200x speedup relies on excluding pre-training cost and a loose 4x4 threshold; the 15x over random curriculum is the honest number. read the letter →

arxiv 2505.00233 v1 pith:OWR6IW7H submitted 2025-05-01 cond-mat.str-el cs.LG

Kimihiro Yamazaki , Takuya Konishi , Yoshinobu Kawahara This is my paper · ORCID
classification cond-mat.str-elcs.LG
keywords CurriculumLearningTransferNeuralNetworkQuantumStatesStronglyCorrelatedElectronSystemsHubbardmodelVariationalMonteCarloPerturbationTheory
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that exploring many different parameter regimes of the Hubbard model with neural-network quantum states can be made hundreds of times cheaper by treating the sequence of training runs as a curriculum: start at one interaction strength, and always fine-tune to the nearest neighboring parameter value next. The justification is physical: a small change in the Hamiltonian parameters is a small perturbation, so the previously trained wave function is a near-optimal starting point, and the zero-shot energy of that starting point exactly matches the first-order perturbation-theory energy. The authors build a transferable network architecture, Pairing-Net, that splits the wave function into a neural pair-product state and a neural correlation factor, and show on 2x2 to 6x6 Hubbard lattices that the nearest-neighbor curriculum converges in a handful of epochs with near-100% success, versus hundreds of epochs and frequent failures when training each parameter from scratch. If right, this makes exhaustive scanning of interaction strength and electron filling feasible for systems where it was previously prohibitively expensive.

What carries the argument

The central mechanism is the identification of NQS fine-tuning with perturbation theory (Eqs. 6-8): the parameter difference ν_k = λ_k - λ_{k-1} generates the perturbation operator V̂_k = Ĥ_k - Ĥ_{k-1}, and the energy of the zero-shot (un-fine-tuned) network equals the first-order perturbation energy. This identity turns the abstract problem of what order to explore parameters into a geometric one: smaller |ν_k| means a smaller perturbation, so the optimal curriculum is the nearest-neighbor path in parameter space (Eq. 9). The load-bearing object that makes this practical is Pairing-Net, a wave-function ansatz Ψ = Pf[X_θ] × exp(C_φ(x)) combining a neural pair-product (Pfaffian) state with a neural correlation factor, whose input structure is independent of λ so the same network can be fine-tuned across interaction strengths and fillings.

What would settle it

Run a Hubbard-model curriculum on a 4x4 lattice along a parameter path that must cross a first-order transition or a level crossing in a small cluster, and compare the total epoch count of the nearest-neighbor path against a path that approaches the target from the same phase; if the same-phase path converges in fewer total epochs, the nearest-neighbor rule is falsified. A cheaper check: measure ΔE^(0) for fine-tuning steps of equal |ν| taken on opposite sides of a known phase boundary — the rule predicts both should need comparably few epochs, while a non-perturbative breakdown would show one side requiring far more.

Watch

Extended reading notes

Core claim

The paper establishes that transfer learning between neural network quantum states for the Hubbard model is quantitatively the same operation as perturbation theory: initializing the wave function at a new parameter value and evaluating its energy gives exactly the first-order Rayleigh-Schrödinger energy correction, and fine-tuning the network numerically renormalizes all higher-order corrections without needing their analytic form. From this, it derives a curriculum rule (Eq. 9): the optimal order to visit a set of parameter points is always to move to the nearest unexplored neighbor, because the perturbation is smallest and the pre-trained state is closest to the target ground state. On the single-axis parameter sets tested (U/t = 2, 4, 6, 8 at half-filling, and electron densities around half-filling), this curriculum converges in roughly 3 to 5 epochs per step on 4x4 lattices, while random curricula take about 45 total epochs and from-scratch runs take about 792, i.e., a 15-fold and 200-fold speedup respectively, with zero variance in convergence epochs versus large variance for conventional runs. The claim is validated for both the pair-product Pairing-Net architecture and an alternative Slater-determinant network, and it persists even when fine-tuning crosses a known quantum phase boundary in the J1-J2 Heisenberg model.

Load-bearing premise

The curriculum rule's ordering assumption — that nearest-neighbor parameter steps are always fastest because smaller perturbations are easier to correct — is derived from perturbation theory and is not guaranteed in strong-coupling (U/t ≳ 6) or near-critical regimes, where the paper relies on empirical evidence rather than proof.

Editorial extensions

If this is right

  • Exhaustive parameter scans of Hubbard-type models (U/t, filling) drop to a small fraction of the epoch cost: for the 4x4 half-filling set, the whole curriculum needs about 3 fine-tuning epochs total versus roughly 792 from scratch.
  • The speedup grows with system size: the curriculum's advantage over random ordering is larger at M=4x4 than at M=2x2, so the method pays off most where conventional NQS-VMC is most expensive.
  • Transfer learning doubles as a numerical higher-order perturbation theory: zero-shot energies give first-order corrections, and fine-tuning renormalizes all higher orders without an analytic expansion.
  • The method is architecture-agnostic (it works with Pfaffian pair-product and Slater-determinant networks) and remains effective across a phase boundary in the J1-J2 Heisenberg model, suggesting broad applicability beyond the Hubbard model.
  • Optimization stability improves dramatically: fine-tuning converged in 1000 of 1000 independent trials with zero standard deviation in epochs, versus as low as a 1.9% convergence rate for from-scratch runs.

Reading between the lines

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

  • The nearest-neighbor rule could be extended to multi-axis sweeps by treating the parameter grid as a graph and walking the shortest Hamiltonian path; the paper only tests single-axis variations, but the perturbative argument suggests the same ordering should guide any path through (Ñ, Ũ) space.
  • The first-order-perturbation interpretation gives a free convergence diagnostic: the zero-shot energy error ΔE^(0) predicts how many fine-tuning epochs a step will need, so an explorer could adaptively choose step sizes to keep ΔE^(0) below a threshold.
  • Because the method is architecture-agnostic, pairing it with more expressive backbones (such as the Vision Transformer wave functions already demonstrated in the appendix) could extend the speedup from the Hubbard model to frustrated spin systems and, with multiorbital inputs, to realistic material models, though this is not tested in the paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a curriculum-learning framework for neural-network quantum states (NQS) aimed at exhaustively exploring the parameter space of strongly correlated electron systems. The curriculum is generated by iteratively fine-tuning a transferable NQS between nearest-neighbor points in physical parameter space, and this ordering is justified through a perturbative reading of transfer learning, where the zero-shot energy of a pre-trained state is shown to be the first-order perturbation energy. The authors introduce Pairing-Net, an architecture combining a neural pair-product state with a neural correlation factor, and test transfer and curriculum learning on 2x2 and 4x4 (with 6x6 in an appendix) Hubbard models and on a J1-J2 Heisenberg model. The main quantitative claims are a roughly 15-fold speedup over a random curriculum and over 200-fold speedup over conventional from-scratch VMC, together with improved convergence stability, as reported in Table 2.

Significance. Assuming the quantitative claims can be made robust, the paper makes a useful contribution: it provides a principled ordering rule for transfer-learning tasks on NQS, a concrete transferable architecture, and positive evidence across two architectures (Pairing-Net and SD-Net) and three models, with ground states checked against exact diagonalization, AF-QMC, and RBM+PP references. The identity in Eq. (8) is exact and needs no fitting, and the comparison between ordered and random curricula is the right kind of empirical test for the ordering rule. However, the headline speedup figures are currently not supported as stated because of inconsistent convergence thresholds, the exclusion of pre-training epochs, and an insufficiently characterized random baseline. These issues are fixable within the manuscript's scope but require re-analysis or substantial re-reporting.

major comments (3)
  1. [§7.3, Table 2 and Appendix E] The headline speedup claim rests on a comparison that is not end-to-end. Table 2 reports "total number of epochs excluding pre-training", yet the conventional (N/A) baseline includes only from-scratch epochs; the pre-training stage costs 1000 epochs per architecture (Table C1, Appendix D), which for the K=4 task set already exceeds the 792-epoch conventional baseline. In addition, the convergence criterion is 1000 times looser at 4x4 than at 2x2 (dE < 1e-1 versus dE < 1e-4, Appendix E), and several 4x4 transfer rows in Table E2 list 0 epochs because the zero-shot energy dE^(0) is already below the threshold. The total of 3 epochs for the 4x4 curriculum is therefore largely a threshold artifact. The claim "over 200-fold speedup in computation" in the abstract should be replaced by a properly defined end-to-end metric, including pre-training and wall-clock time, or explicitly restricted to fine-tuning epochs under a consistent convergence criterion across system sizes.
  2. [§7.3, Table 2 and Fig. E2] The 'Random' entry in Table 2 is not defined as a statistic over random curricula. Figure E2 refers to a "randomly selected poor curriculum", which suggests that the 45-epoch number is a single unfavorable instance rather than an average; a single instance cannot support the claim of "up to a 15-fold speedup compared to a random curriculum". Moreover, the comparison of "Ours" versus "N/A" conflates two effects: the benefit of any transfer (a better initial wavefunction) and the benefit of the curriculum order. Because the conventional baseline is trained from scratch while all curriculum variants inherit a pre-trained state, the 200-fold factor cannot be attributed to the curriculum rule. The authors should report the mean and spread over all (or many) random permutations, and should decompose the epoch counts to separate initialization advantage from ordering advantage.
  3. [§5.3 and Appendix I] The curriculum rule in Eq. (9) is derived under a perturbative assumption, and the authors themselves note that it is not obvious the rule holds in strong-coupling or near-critical regimes. The empirical evidence in Appendix I shows that individual fine-tuning steps remain effective when crossing J2/J1 ~ 0.49, but it does not compare nearest-neighbor ordering against alternative orderings in that regime. Section 7.3's comparison is only for the Hubbard task set U = 2, 4, 6, 8, which does not include a phase boundary. To support the claim that the nearest-neighbor rule is optimal in strongly correlated or near-critical settings, the authors should report a randomized-order comparison in at least one such regime, or temper the claim in Section 7.1 that "these results strongly suggest that the rule remains valid in such scenarios".
minor comments (6)
  1. [§7.1] The first paragraph contains a duplicated sentence: "Here, we present benchmark results demonstrating that transfer learning with Pairing-Net improves both computational efficiency and stability" appears twice; one copy should be removed.
  2. [Appendix E.2] The sentence "Table E2, upper panel, summarizes the results ... with respect to electron density" is incorrect: the upper panel shows transfer in U at fixed density, while the lower panel shows transfer in electron density. The cross-reference should be fixed.
  3. [Abstract and Table 2] The phrase "approximately 200-fold speedup in computation" in the abstract should be qualified, since Table 2 reports epochs excluding pre-training and uses different convergence thresholds for 2x2 and 4x4; as written, it overstates the computational saving.
  4. [Reproducibility statement] The statement "The implementation code used to generate all results in this paper is available from the authors upon reasonable request" is not adequate for a computational methods paper; depositing the code in a public repository would allow the numerical claims to be verified.
  5. [Eq. (9)] The indexing in Eq. (9) is ambiguous at the endpoints of the parameter sequence: the condition involving k-1 and k+1 is not defined for k=1 or k=K. Please clarify how the nearest-neighbor path is selected, including the final step.
  6. [Fig. E2] The caption says the right panel shows "a randomly selected poor curriculum", but Table 2's 'Random' entry is not explicitly tied to this instance; the figure and table should be consistent and should state whether the reported random number is one sample or an average.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the perturbation-theory interpretation is an algebraic identity and the curriculum rule is empirically validated against external benchmarks.

full rationale

The paper's derivation chain contains no step in which an input is renamed as a prediction or defined in terms of the claimed output. Equation (8), identifying the zero-shot energy with the unperturbed energy plus the first-order perturbation correction, is an exact algebraic identity: E_{\theta^{(0)}_k} = \langle\Psi_{\theta_{k-1}}|\hat{H}_k|\Psi_{\theta_{k-1}}\rangle/\langle\Psi_{\theta_{k-1}}|\Psi_{\theta_{k-1}}\rangle = E_{\theta_{k-1}} + \langle\Psi_{\theta_{k-1}}|\hat{V}_k|\Psi_{\theta_{k-1}}\rangle/\langle\Psi_{\theta_{k-1}}|\Psi_{\theta_{k-1}}\rangle. This is an interpretation of transfer learning in perturbative language, not a fitted quantity disguised as a finding. The curriculum rule in Eq. (9) is derived heuristically from perturbation theory and then tested empirically in Section 7.3 and Appendices E.4 and I; its validity is not assumed by the speedup claim, and the random-curriculum baseline provides an independent comparison. All central accuracy claims are benchmarked against exact diagonalization (M = 2x2), AF-QMC (M = 4x4), and RBM+PP reference results (J1-J2 model), so the measured gains are not manufactured by construction. No load-bearing self-citation appears: prior work is cited for context and baselines, not to justify the proposed curriculum rule. The skeptical concerns about the 15x/200x speedup figures concern the fairness of the comparison (excluded pre-training epochs, a looser 4x4 convergence threshold, and a from-scratch baseline that does not inherit a good initialization). Those are correctness or reporting concerns, not circularity, because the central qualitative finding, that nearest-neighbor transfer curricula reduce epoch counts relative to random curricula, is supported by the directly measured per-task epoch data in Table E2 and Figure E2.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on empirical benchmarks rather than a closed-form derivation, so the ledger contains the benchmark reference assumptions, the hand-chosen thresholds and epoch budgets that shape the speedup numbers, and the ordering rule from Eq. (9). No new physical entities are introduced; Pairing-Net is a neural-network architecture, not a postulated physical object.

free parameters (3)
  • Convergence threshold epsilon = 10^-4 for 2x2, 10^-1 for 4x4
    Defines convergence and directly controls reported epochs-to-convergence; with 10^-1, several transfer runs count as 0 epochs, inflating the speedup. Location: Section 7 and Appendix E.
  • Pre-training epoch count = 1000
    Fixed for all systems and excluded from curriculum epoch totals in Table 2; this exclusion underlies the 200-fold claim. Location: Table C1.
  • Architecture hyperparameters = NPP hidden 50-100, NCF hidden 100-400, learning rate 0.05 to 0.01
    Chosen by hand without a sensitivity analysis; they affect both accuracy and convergence speed. Location: Table C1 and Appendix D.
assumptions (6)
  • domain assumption VMC with stochastic reconfiguration converges to the ground state within a small energy tolerance for the chosen trial ansatz.
    The entire empirical method relies on optimization finding the ground-state NQS; used throughout Section 7.
  • domain assumption AF-QMC energies are accurate ground-state references for 4x4 and 6x6 Hubbard models.
    Used as Eg in Delta E for M = 4x4 and M = 6x6; cited as [19]. If these references are biased, all error metrics shift.
  • domain assumption RBM+PP energies are accurate references for the 6x6 J1-J2 Heisenberg model.
    Appendix I uses the results of Nomura and Imada [40] as ground truth.
  • ad hoc to paper Small parameter changes induce small perturbation operators, so nearest-neighbor fine-tuning is optimal.
    This is the paper's curriculum rule, Eq. (9). It is argued from perturbation theory but relied on empirically in strong-coupling and near-critical cases; it is not proven.
  • domain assumption Pairing-Net and SD-Net ansatze are sufficiently expressive to represent the ground states of the studied systems.
    Empirically supported by benchmark energies, but not guaranteed; if expressivity fails, convergence times are not representative.
  • standard math The Pfaffian of a skew-symmetric matrix yields an antisymmetric wavefunction.
    Used in Appendix C to enforce fermionic antisymmetry for Pairing-Net.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explorative Curriculum Learning for Strongly Correlated Electron Systems." pith.science (2026). https://pith.science/paper/OWR6IW7H

@misc{pith2026250500233,
  author       = {Pith},
  title        = {Pith review of: Explorative Curriculum Learning for Strongly Correlated Electron Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWR6IW7H}},
  note         = {Machine review of arXiv:2505.00233}
}
read the original abstract

Recent advances in neural network quantum states (NQS) have enabled high-accuracy predictions for complex quantum many-body systems such as strongly correlated electron systems. However, the computational cost remains prohibitive, making exploration of the diverse parameters of interaction strengths and other physical parameters inefficient. While transfer learning has been proposed to mitigate this challenge, achieving generalization to large-scale systems and diverse parameter regimes remains difficult. To address this limitation, we propose a novel curriculum learning framework based on transfer learning for NQS. This facilitates efficient and stable exploration across a vast parameter space of quantum many-body systems. In addition, by interpreting NQS transfer learning through a perturbative lens, we demonstrate how prior physical knowledge can be flexibly incorporated into the curriculum learning process. We also propose Pairing-Net, an architecture to practically implement this strategy for strongly correlated electron systems, and empirically verify its effectiveness. Our results show an approximately 200-fold speedup in computation and a marked improvement in optimization stability compared to conventional methods.

Figures

Figures reproduced from arXiv: 2505.00233 by the authors.

Figure 1
Figure 1. Illustration of the parameter order in curriculum learning. Circles and the filled one represent possible targets and the pre-training task, respectively, in the parameter space of λ. This illustrates that iteratively performing transfer learning to the nearest neighboring parameter in the space is crucial for efficient and stable curriculum learning. criterion with the L1 norm properly determines the task order. We… view at source ↗
Figure 2
Figure 2. Pairing-Net architecture. The input to the network is generated from a single sampled configuration, and the output is the wave function amplitude Ψθk,φk (x), parameterized by neural networks θk and φk. Pairing-Net decomposes the wave function into two neural networks: A neural pair-product (NPP) wave function Ψ(NPP) θk (x) and a neural correlation factor (NCF) Cφk (x). The input to the NPP wave function consists of… view at source ↗
Figure 3
Figure 3. The loss curves for M = 4×4 over 100 epochs applied to (N˜ 2,U˜ 2) = (16/16, 4) using the various pre-trained NQS. maximum of 500 epochs, is also shown. The mean number of epochs to convergence serves as a metric for efficiency, while the std. of the number of epochs to convergence and the convergence count are used as metrics for stability. The results demonstrate that transfer learning achieves significant computa… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The dependence of ∆E(0), acquired through pre-training, on changes in the target parameter U˜ 2. Pre-trained NQS with U˜ 1 = 2, 4, 6, 8 were used. remarkably, these results strongly suggest that the rule remains valid in such scenarios. This robustness can be attribute…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 26 canonical work pages

  1. [1]

    A quantum many-body system consists of multiple interacting particles, and its properties are governed by the quantum state

    Introduction Obtaining quantum states of various quantum many-body systems is one of the most important problems in physics. A quantum many-body system consists of multiple interacting particles, and its properties are governed by the quantum state. Among these systems, strongly correlated electron systems are known to be particularly difficult to calcula...

    work page

  2. [2]

    Background 2.1. Quantum Many-Body Problem and Strongly Correlated Electron Systems The quantum many-body problem, meaning here the problem of obtaining quantum states in a quantum many-body system, is essentially reduced to solving the following eigenvalue problem for a Hamiltonian operator ˆH : H→ H: ˆH|Ψ⟩ =E|Ψ⟩, (1) where H is a Hilbert space, |Ψ⟩∈ H is...

    work page

  3. [3]

    Related Works Conventional VMC-based NQS methods have been hampered by high computational costs and numerical instability. While recent works have demonstrated the potential of transfer learning to mitigate these issues across various quantum many-body systems [7, 8, 10, 11, 12], the effectiveness of transfer learning in this context remains limited. More...

    work page

  4. [4]

    Curriculum Learning for Quantum Many-Body Problem Here we propose a novel curriculum learning framework aimed at improving the efficiency and stability of exploration of NQS across diverse parameters for quantum many-body systems. 4.1. Transfer Learning with NQS We formalize the transfer learning of NQS, a crucial component of our proposed method. Convent...

    work page

  5. [5]

    renormalization

    Perturbation Theory Perspective This section demonstrates that the above transfer learning (or curriculum learning with K = 2) can be interpreted as perturbation theory. This interpretation reveals that transfer learning constitutes a rigorous numerical method grounded in the fundamental principles of renormalization in theoretical physics. Furthermore, w...

    work page

  6. [6]

    neural pair product (NPP) state

    Pairing-Net In this section, we introduce Pairing-Net, a transferable NQS architecture capable of performing transfer and curriculum learning for strongly correlated electron systems. Existing approaches require independent training of NQS for each of the K tasks in the given task set{T (λk)}K k=1. For the Hubbard model, this corresponds to exploring K di...

    work page

  7. [7]

    First, we demonstrate that transfer learning using Pairing-Net improves computational efficiency and stability

    Experimental Results This section presents several numerical experiments demonstrating the effectiveness of our method, including benchmark evaluations and practical applications for two- dimensional Hubbard models on a square lattice with periodic boundary conditions. First, we demonstrate that transfer learning using Pairing-Net improves computational e...

    work page

  8. [8]

    The framework introduces transfer learning for NQS and subsequently proposes a novel curriculum learning algorithm based on iterative transfer learning

    Conclusions In this study, we propose a novel curriculum learning framework for the efficient and stable exploration of a broad parameter space in quantum many-body systems. The framework introduces transfer learning for NQS and subsequently proposes a novel curriculum learning algorithm based on iterative transfer learning. By recognizing the interpretab...

    work page arXiv

Show all 48 references
  1. [9]

    Imada M, Fujimori A and Tokura Y 1998 Rev. Mod. Phys. 70(4) 1039–1263 URL https: //link.aps.org/doi/10.1103/RevModPhys.70.1039

    1998 doi

  2. [10]

    Hirsch J E 1985 Phys. Rev. B 31(7) 4403–4419 URL https://link.aps.org/doi/10.1103/ PhysRevB.31.4403

    1985

  3. [11]

    Anderson P W 1997 The Theory of Superconductivity in the High-T”c” Cuprate Superconductors (Princeton University Press) URL http://www.jstor.org/stable/j.ctt1m323t3

    1997

  4. [12]

    Series A

    Hubbard J and Flowers B H 1963Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences 276 238–257 URL https://royalsocietypublishing.org/doi/abs/10. 1098/rspa.1963.0204

    1963

  5. [13]

    1126/science.aag2302

    Carleo G and Troyer M 2017 Science 355 602–606 URL https://www.science.org/doi/abs/10. 1126/science.aag2302

    2017

  6. [14]

    Nomura Y, Darmawan A S, Yamaji Y and Imada M 2017 Phys. Rev. B 96(20) 205152 URL https://link.aps.org/doi/10.1103/PhysRevB.96.205152

    2017 doi

  7. [15]

    mlr.press/v202/gao23c.html

    Gao N and G¨ unnemann S 2023 Proceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research vol 202) ed Krause A, Brunskill E, Cho K, Engelhardt B, Sabato S and Scarlett J (PMLR) pp 10708–10726 URL https://proceedings. mlr.press/...

    2023

  8. [16]

    Scherbela M, Gerard L and Grohs P 2023 Thirty-seventh Conference on Neural Information Processing Systems URL https://openreview.net/forum?id=FBNyccPfAu

    2023

  9. [17]

    Zhu Z, Mattheakis M, Pan W and Kaxiras E 2023 Phys. Rev. Res. 5(4) 043084 URL https: //link.aps.org/doi/10.1103/PhysRevResearch.5.043084

    2023 doi

  10. [18]

    org/10.1038/s41467-023-44216-9

    Scherbela M, Gerard L and Grohs P 2024 Nature Communications 15 120 URL https://doi. org/10.1038/s41467-023-44216-9

    2024 doi

  11. [19]

    Kim J, Pescia G, Fore B, Nys J, Carleo G, Gandolfi S, Hjorth-Jensen M and Lovato A 2024 Communications Physics 7 148 URL https://doi.org/10.1038/s42005-024-01613-w

    2024 doi

  12. [20]

    Gao N and G¨ unnemann S 2024 The Thirty-eighth Annual Conference on Neural Information Processing Systems URL https://openreview.net/forum?id=HRkniCWM3E

    2024

  13. [21]

    Rende R, Goldt S, Becca F and Viteritti L L 2024 Phys. Rev. Res. 6(4) 043280 URL https: //link.aps.org/doi/10.1103/PhysRevResearch.6.043280

    2024 doi

  14. [22]

    Sakakibara H and Kotani T 2019 Phys. Rev. B 99(19) 195141 URL https://link.aps.org/doi/ 10.1103/PhysRevB.99.195141

    2019 doi

  15. [23]

    Rut G, Fidrysiak M, Goc-Jag lo D and Rycerz A 2023 International Journal of Molecular Sciences 24 URL https://www.mdpi.com/1422-0067/24/2/1509

    2023

  16. [24]

    Iimura S, Matsuishi S, Sato H, Hanna T, Muraba Y, Kim S W, Kim J E, Takata M and Hosono H 2012 Nature communications 3 943 URL https://doi.org/10.1038/ncomms1913

    2012 doi

  17. [25]

    Suzuki K, Usui H, Iimura S, Sato Y, Matsuishi S, Hosono H and Kuroki K 2014 Phys. Rev. Lett. 113(2) 027002 URL https://link.aps.org/doi/10.1103/PhysRevLett.113.027002

    2014 doi

  18. [26]

    McMillan W L 1965 Phys. Rev. 138(2A) A442–A451 URL https://link.aps.org/doi/10.1103/ PhysRev.138.A442

    1965

  19. [27]

    Qin M, Shi H and Zhang S 2016 Phys. Rev. B 94(8) 085103 URL https://link.aps.org/doi/ 10.1103/PhysRevB.94.085103

    2016 doi

  20. [28]

    Amari S i 1998 Neural Computation 10 251–276 URL https://doi.org/10.1162/ 089976698300017746

    1998

  21. [29]

    Sorella S 2001 Phys. Rev. B 64(2) 024512 URL https://link.aps.org/doi/10.1103/PhysRevB. 64.024512

    2001 doi

  22. [30]

    Hastings W K 1970 Biometrika 57 97–109 URL https://doi.org/10.1093/biomet/57.1.97

    1970 doi

  23. [31]

    Wu D, Rossi R, Vicentini F, Astrakhantsev N, Becca F, Cao X, Carrasquilla J, Ferrari F, Georges A, Hibat-Allah M, Imada M, L¨ auchli A M, Mazzola G, Mezzacapo A, Millis A, Moreno J R, Neupert T, Nomura Y, Nys J, Parcollet O, Pohle R, Romero I, Schmid M, Silvester J M, Sorella ...

    2024 doi

  24. [32]

    Bengio Y, Louradour J, Collobert R and Weston J 2009Proceedings of the 26th Annual International Conference on Machine Learning ICML ’09 (New York, NY, USA: Association for Computing Machinery) pp 41–48 URL https://doi.org/10.1145/1553374.1553380

  25. [33]

    Curry R, Lynn J E, Schmidt K E and Gezerlis A 2023 Phys. Rev. Res. 5(4) L042021 URL https://link.aps.org/doi/10.1103/PhysRevResearch.5.L042021

    2023 doi

  26. [34]

    Pfau D, Spencer J S, Matthews A G D G and Foulkes W M C 2020 Phys. Rev. Res. 2(3) 033429 URL https://link.aps.org/doi/10.1103/PhysRevResearch.2.033429

    2020 doi

  27. [35]

    1038/s41557-020-0544-y

    Hermann J, Sch¨ atzle Z and No´ e F 2020Nature Chemistry 12 891–897 URL https://doi.org/10. 1038/s41557-020-0544-y

  28. [36]

    sciencedirect.com/science/article/pii/0375947471902673

    Gaudin M, Gillespie J and Ripka G 1971 Nuclear Physics A 176 237–260 URL https://www. sciencedirect.com/science/article/pii/0375947471902673

    1971

  29. [37]

    Viteritti L L, Rende R and Becca F 2023 Phys. Rev. Lett. 130(23) 236401 URL https: //link.aps.org/doi/10.1103/PhysRevLett.130.236401

    2023 doi

  30. [38]

    Rende R, Viteritti L L, Bardone L, Becca F and Goldt S 2024 Communications Physics 7 260 URL https://doi.org/10.1038/s42005-024-01732-4

    2024 doi

  31. [39]

    Luo D and Clark B K 2019 Phys. Rev. Lett. 122(22) 226401 URL https://link.aps.org/doi/ 10.1103/PhysRevLett.122.226401

    2019 doi

  32. [40]

    Arovas D P, Berg E, Kivelson S A and Raghu S 2022 Annual Review of Condensed Matter Physics 13 239–274 URL https://www.annualreviews.org/content/journals/10.1146/ annurev-conmatphys-031620-102024

    2022

  33. [41]

    Kuroki K, Onari S, Arita R, Usui H, Tanaka Y, Kontani H and Aoki H 2008 Phys. Rev. Lett. 101(8) 087004 URL https://link.aps.org/doi/10.1103/PhysRevLett.101.087004

    2008 doi

  34. [42]

    Maier T, Berlijn T and Scalapino D J 2019 Phys. Rev. B 99(22) 224515 URL https://link.aps. org/doi/10.1103/PhysRevB.99.224515

    2019 doi

  35. [43]

    Bohm D and Pines D 1951 Phys. Rev. 82(5) 625–634 URL https://link.aps.org/doi/10.1103/ PhysRev.82.625

    1951

  36. [44]

    Bickers N E, Scalapino D J and White S R 1989 Phys. Rev. Lett. 62(8) 961–964 URL https://link.aps.org/doi/10.1103/PhysRevLett.62.961

    1989 doi

  37. [45]

    Codebases 7 URL https: //scipost.org/10.21468/SciPostPhysCodeb.7

    Vicentini F, Hofmann D, Szab´ o A, Wu D, Roth C, Giuliani C, Pescia G, Nys J, Vargas- Calder´ on V, Astrakhantsev N and Carleo G 2022 SciPost Phys. Codebases 7 URL https: //scipost.org/10.21468/SciPostPhysCodeb.7

    2022 doi

  38. [46]

    Rende R and Loris Viteritti L 2025 Machine Learning: Science and Technology 6 010501 URL https://dx.doi.org/10.1088/2632-2153/ada1a0

    2025 doi

  39. [47]

    Hu W J, Becca F, Parola A and Sorella S 2013 Phys. Rev. B 88(6) 060402 URL https: //link.aps.org/doi/10.1103/PhysRevB.88.060402

    2013 doi

  40. [48]

    Nomura Y and Imada M 2021 Phys. Rev. X 11(3) 031034 URL https://link.aps.org/doi/10. 1103/PhysRevX.11.031034

    2021

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.