Pith. sign in

REVIEW 4 minor 53 references

Diagonal Symmetrization of Neural Network Solvers for the Many-Electron Schr\"odinger Equation

T0 review · 0 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read In-training symmetrization of neural wavefunctions inflates gradient variance under fixed compute, while post hoc averaging of the trained wavefunction improves energy, variance, and symmetry.

desk verdict Careful equal-compute analysis shows post hoc averaging is a cheap win and in-training symmetrization can hurt—but the negative claim is scoped to a cost regime the abstract overstates. read the letter →

arxiv 2502.05318 v2 pith:MUN4JYPW submitted 2025-02-07 cs.LG cond-mat.mtrl-sci

Kevin Han Huang , Ni Zhan , Elif Ertekin , Peter Orbanz , Ryan P. Adams This is my paper · ORCID
classification cs.LGcond-mat.mtrl-sci MSC 81V5581V7065C0568T07
keywords variationalMonteCarloneuralwavefunctionansatzdiagonalgroupsymmetryspacedataaugmentationaveragingposthocsymmetrizationcomputational-statisticaltradeoff
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

This paper asks whether enforcing the diagonal isometry symmetries of a crystal—transformations that move every electron by the same rotation or translation—helps or hurts neural-network solvers for the many-electron Schrödinger equation. The answer it argues for is split: during training, symmetrization under a fixed compute budget inflates the variance of the gradient estimate and can degrade the final energy, contrary to the usual ML expectation that symmetry helps; after training, averaging the learned wavefunction over group elements reliably improves energy, local-energy variance, and symmetry. The key is an accounting rule: each symmetry operation is charged one gradient evaluation, so in-training symmetrization silently replaces independent Monte Carlo samples with transformed copies. On LiH, post hoc averaging reaches -8.1507 Ha versus -8.138(2) Ha for the unsymmetrized baseline, and on graphene -76.050(3) Ha versus -76.039(6) Ha, with PA sometimes matching models trained with ten times the budget. The paper's point matters because VMC solvers are expensive, and the result separates genuine symmetry benefits from the extra compute that symmetrization consumes.

What carries the argument

The central object is the diagonal group $G_{\mathrm{diag}}$ of isometries acting on all $n$ electron positions simultaneously, together with the VMC budget identity that one gradient evaluation dominates one sampling step, so a $k$-fold symmetrization at equal cost must cut the independent-sample count from $N$ to $N/k$. Proposition 4.1 and Lemma 4.2 quantify the resulting variance inflation: for data augmentation, an extra conditional-variance term $(k-1)/N \, \mathrm{Var}\,\mathbb{E}[F_{gX;\psi_\theta}|X]$ appears; for group averaging, the variance is $\mathrm{Var}[F_{X^G;\psi^G_\theta}]/(N/k)$. At inference, Fact 2.1 licenses post hoc averaging over any finite subgroup, and the paper's smoothed canonicalization construction shows why canonicalization is computationally prohibitive in training and numerically unstable post hoc due to second-derivative blowup.

What would settle it

Repeat the graphene experiments under a regime where MCMC sampling cost exceeds gradient cost (longer chains, cheaper gradient evaluations), so the $k$-fold symmetry operations no longer force a reduction from $N$ to $N/k$ samples; the paper's Appendix I predicts the variance inflation of data augmentation disappears, which would falsify the general claim that in-training symmetrization destabilizes VMC.

Watch

Extended reading notes

Core claim

The paper's central claim is that, in variational Monte Carlo training of neural wavefunctions, diagonal-group symmetrization is subject to a computational-statistical tradeoff that standard ML analyses miss. Holding the per-step computational budget fixed, data augmentation and group averaging reduce the number of independent Monte Carlo samples from $N$ to $N/k$ for a $k$-fold symmetry, and this sample reduction inflates the variance of the gradient update: Proposition 4.1 shows that under an invariant chain distribution the data-augmented update has the same expectation as the original update but larger variance by a positive semidefinite term, and Lemma 4.2 shows group averaging scales variance by a factor of $k$. Consequently, in-training symmetrization can destabilize training and yield worse final energies. In contrast, post hoc averaging of the trained wavefunction over a finite subset of the symmetry group—valid because Fact 2.1 ensures such an average remains an eigenfunction with the same energy—improves energy, local-energy variance, and symmetry, with performance sometimes comparable to ten times the training budget.

Load-bearing premise

The load-bearing premise is that gradient evaluation dominates sampling cost, so under a fixed budget symmetrization must reduce independent samples from $N$ to $N/k$; if sampling dominates instead, the paper's own Appendix I concedes data augmentation no longer destabilizes.

Editorial extensions

If this is right

  • Data augmentation and group averaging at equal total compute can degrade or fail to improve the final ground-state energy, even when the symmetry group is large and physically relevant.
  • Post hoc averaging over a finite subgroup improves energy, local-energy variance, and formal symmetry, with benefit growing with subgroup size until saturation (graphene: $-76.050(3)$ Ha; LiH: $-8.1507$ Ha).
  • Post hoc averaged DeepSolid can match the energy of a model trained with roughly ten times the computational budget, making symmetry a cheaper route to accuracy than more training.
  • The computational-statistical tradeoff should transfer to other settings where fresh samples are drawn between gradient updates, such as contrastive divergence training of energy-based models.
  • Smoothed canonicalization is unsuitable for training and also fails post hoc because boundary smoothing inflates second derivatives entering the local energy.

Reading between the lines

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

  • If gradient evaluation later becomes cheaper than sampling (for example through specialized derivative methods), the in-training verdict changes: data augmentation at full sample count no longer inflates variance, per the paper's Appendix I, so the recommendation becomes implementation-dependent.
  • The variance-inflation mechanism suggests a re-reading of earlier reports that equivariant architectures improve energy: part of the gain may be the extra compute consumed by symmetry operations, and only equal-budget comparisons can separate inductive-bias gains from compute gains.
  • Post hoc averaging over strategically chosen generators (not the full group) showed mixed results in the paper; a selection heuristic for which symmetries matter most would be a natural, testable next step.
  • Since post hoc averaging is architecture-agnostic, applying it to any converged neural wavefunction, including molecular FermiNet-style solvers without spatial symmetry, is a cheap and directly testable extension the paper leaves implicit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 4 minor

Summary. The paper studies how to incorporate diagonal group symmetries (space-group isometries acting simultaneously on all electrons) into neural-network variational Monte Carlo solvers for the many-electron Schrödinger equation. Using DeepSolid as the base architecture, it compares three in-training symmetrization strategies—data augmentation (DA), group averaging (GA), and smoothed canonicalization (SC)—against post hoc averaging (PA) of an already-trained wavefunction. The central theoretical results are Proposition 4.1 and Lemma 4.2, which show that, under a fixed per-step computational budget in which gradient cost dominates sampling cost, in-training symmetrization reduces the number of independent MCMC samples from N to N/k and can inflate gradient variance. Empirically, Table 2 and Figure 5 show that DA and GA do not improve or slightly worsen energies on graphene, LiH, and bcc-Li, while PA systematically lowers both energy and local-energy variance and improves symmetry. The paper also develops a smoothed canonicalization adapted to diagonal space-group actions with anti-symmetry preservation, and it documents the numerical cost model in Table 1.

Significance. If the results hold, the paper makes a valuable and somewhat counterintuitive contribution: it identifies a computational-statistical tradeoff specific to VMC, where symmetry augmentation during training is not the free lunch it often is in supervised learning, while inference-time averaging is a cheap and effective alternative. The theoretical derivations are self-contained and carefully stated: Proposition 4.1 and Lemma 4.2 are proved from the stated VMC update rules, the high-dimensional CLT in Appendix D is a nontrivial addition, and the smoothed canonicalization construction in Appendix E comes with explicit invariance, anti-symmetry, and smoothness guarantees. The empirical PA gains are consistent across three distinct systems and reach chemical-accuracy-scale energy differences (e.g., LiH: -8.1507(8) Ha vs -8.138(2) Ha for the baseline; graphene: -76.050(3) vs -76.039(6)), and the released code supports reproducibility.

minor comments (4)
  1. [Abstract and Section 3/Appendix I] The abstract states without qualification that in-training symmetrization 'destabilizes training and can lead to worse performance'; however, Proposition 4.1 and Lemma 4.2 are derived under the fixed-compute model of Section 3, where Csamp << Cgrad forces the N-to-N/k sample reduction, and Appendix I explicitly concedes that when N'=N independent samples can be kept, the DA instability term vanishes. I recommend adding the cost-regime qualifier to the abstract and to the first sentence of Section 4, since the current wording overstates the domain of the negative result.
  2. [Section 7, graphene paragraph and Table 2] The text says that PA 'outperforms other methods both in terms of the metrics in Table 2 and speed of convergence,' but the graphene energy for GAs, -76.049(3) Ha, is statistically indistinguishable from the PA energy -76.050(3) Ha; the claim is better supported by the lower local-energy variance (0.33(1) vs 0.48(2)) and perfect symmetry, so the sentence should be softened or explicitly keyed to those metrics.
  3. [Appendix G.1, proof of Proposition 4.1] In the law-of-total-covariance step, the conditional covariance term is set to zero because g1,1 and g1,2 are independent conditioned on X1; adding one sentence to that effect would make the proof easier to follow.
  4. [Section 6, notation Var[PA/OG]] The notation Var[PA/OG] is introduced only in Table 2 and defined in Section 6; for readers who start from the table, a parenthetical clarification in the caption or a pointer to Section 6 would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical results are self-contained proofs from stated VMC cost assumptions, and the post hoc averaging gains are validated on energies not used to fit any parameter.

full rationale

The paper's derivation chain is self-contained rather than circular. Section 3 defines the VMC setup and explicitly states the cost model Csamp << Cgrad; Section 4 then defines the original, data-augmented, and group-averaged updates with concrete batch sizes, and Proposition 4.1 and Lemma 4.2 derive their mean and variance by standard i.i.d. sample-size calculus and the law of total covariance. The DA variance inflation follows mathematically from reducing independent MCMC draws from N to N/k under a fixed gradient budget; it is not obtained by fitting a parameter and then relabeling the fit as a prediction. The claimed tradeoff is conditional on the stated cost model, and Appendix I explicitly concedes that if Csamp >= Cgrad and one keeps N' = N independent samples, DA does not destabilize. That is an honest scope limitation, not a circular step. Fact 2.1 is a classical linearity argument for the Schrödinger equation, and the post hoc averaging results are evaluated on energies and local-energy variances that were not used to fit the wavefunction parameters. The self-citations to Huang et al. (2022) are background context for standard ML data-augmentation analyses and for a CLT statement, but they are not load-bearing: Theorem D.1 is derived from external results in Chernozhukov et al. (2017), and the central variance propositions are proven in the appendix from the paper's own definitions. No uniqueness theorem or invariant-map ansatz is imported from the authors' prior work to force the paper's conclusions. Overall, the central claim has independent mathematical and numerical content, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central PA result rests on domain assumptions about VMC and on a fixed compute budget, not on fitted parameters. The theoretical variance statements depend on the i.i.d.-chain idealization and the C_samp << C_grad model. No ad hoc fitted constants are introduced to obtain the main results, and no new physical entities are postulated.

free parameters (2)
  • epsilon (smoothing bandwidth for smoothed canonicalization) = not specified precisely (Figure 10 illustrates 0.1)
    Hand-chosen in Appendix E; controls boundary averaging region and derivative blowup in SC/PC. Not fitted to energy data and does not affect the PA central result.
  • smoothing step function s_inf or s_2 = s_inf and s_2 variants
    Differ in smoothness and derivative blowup; PC results in Table 2 use s_inf. Hand choice, not tuned; relevant only to the negative canonicalization results.
assumptions (6)
  • domain assumption Born-Oppenheimer approximation: nuclei are fixed classical charges r_I and the electronic Hamiltonian is H = -1/2 Delta + V_Coul.
    Stated in Section 1.1 and used throughout; the entire VMC setting and diagonal invariance of V rely on it.
  • domain assumption The space group G_sp can be truncated to a finite group by the supercell assumption, so Fact 2.1 applies to finite averaging sets.
    Section 2: the supercell restriction effectively reduces the translations to a finite set, making the group finite.
  • domain assumption MCMC samples from chains of length m are treated as i.i.d. draws from the chain distribution p_{psi_theta}^{(m)} in the theoretical analysis.
    Sections 3 and 4 define X_1,...,X_N i.i.d. from p^{(m)}; real chains are correlated, which is not modeled in Proposition 4.1.
  • domain assumption Computational cost is dominated by gradient evaluation (C_samp << C_grad), so equal-compute comparisons reduce the number of fresh samples by the symmetry order k.
    Section 3 and the DA/GA definitions in Section 4; Appendix I explicitly notes the alternative regime. This is the weakest premise.
  • standard math The high-dimensional CLT of Chernozhukov, Chetverikov, and Kato (2017) and the Berry-Esseen theorem are assumed to characterize gradient update distributions.
    Theorem D.1 is built on these cited results; they are not re-proved in the paper.
  • domain assumption DeepSolid's architecture and hyperparameters from Li et al. (2022) are taken as a fixed, representative VMC solver.
    Section 7 and Appendix B keep DeepSolid defaults; generality beyond this architecture is conjectured in Section 8.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diagonal Symmetrization of Neural Network Solvers for the Many-Electron Schr\"odinger Equation." pith.science (2026). https://pith.science/paper/MUN4JYPW

@misc{pith2026250205318,
  author       = {Pith},
  title        = {Pith review of: Diagonal Symmetrization of Neural Network Solvers for the Many-Electron Schr\"odinger Equation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MUN4JYPW}},
  note         = {Machine review of arXiv:2502.05318}
}
read the original abstract

Incorporating group symmetries into neural networks has been a cornerstone of success in many AI-for-science applications. Diagonal groups of isometries, which describe the invariance under a simultaneous movement of multiple objects, arise naturally in many-body quantum problems. Despite their importance, diagonal groups have received relatively little attention, as they lack a natural choice of invariant maps except in special cases. We study different ways of incorporating diagonal invariance in neural network ans\"atze trained via variational Monte Carlo methods, and consider specifically data augmentation, group averaging and canonicalization. We show that, contrary to standard ML setups, in-training symmetrization destabilizes training and can lead to worse performance. Our theoretical and numerical results indicate that this unexpected behavior may arise from a unique computational-statistical tradeoff not found in standard ML analyses of symmetrization. Meanwhile, we demonstrate that post hoc averaging is less sensitive to such tradeoffs and emerges as a simple, flexible and effective method for improving neural network solvers.

Figures

Figures reproduced from arXiv: 2502.05318 by the authors.

Figure 1
Figure 1. Visualizations of the (partial) diagonal invariance of an unsymmetrized wavefunction versus a symmetrized wavefunction in a graphene 1 × 1 system. (a) and (b) are generated by evaluating log |ψ(x1 + t, . . . , x12 + t)| 2 under a simultaneous 2d translation t of the configuration (x1, . . . , x12) given by the 12 blue spheres in (e). The red overlay indicates the unit cell in (e) such that the ref. line is exactly a… view at source ↗
Figure 2
Figure 2. Different invariances for two electrons in a graphene system. Left. Separate invariance under a reflection and a rotation. Right. Diagonal invariance under a simultaneous reflection. empirical results in the special case of a smoothed canoni￾calization (Sec. 4.3 and Appendix E). Sec. 3 briefly reviews the VMC setup and how different computational costs arise. Sec. 4 and 5 respectively examine in-training and post ho… view at source ↗
Figure 3
Figure 3. Normalized variance of differently symmetrized gradient updates against GPU hours. Experiment details in Sec. 7. Method N k Csamp (s) Cgrad (s) Total (s) OG 1000 - 0.16(3) 2.4(3) 2.5(3) DA 90 12 0.041(5) 2.4(2) 2.5(2) GA 90 12 0.16(2) 2.6(1) 2.7(1) GA 1000 12 1.50(1) 24(1) 25(1) GAs 1000 1 0.16(1) 2.4(1) 2.5(1) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Canonicalization functions for 1d unit translations. under KFAC, a second-order method used by DeepSolid, is also inflated under DA [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Performance of wavefunctions against GPU hours, ob￾tained with different symmetrization methods. Metrics are defined in Sec. 6 and experiment details in Sec. 7. (c) is computed via Var[PA/OG] in Sec. 6 but with ψ (OG) θ replaced by different ψθ’s. sical wavefunction an…
Figure 6
Figure 6. Figure 6: Visualization of the primitive cells of LiH and bcc-Li. The experimental benchmark used for comparing all symmetrization strategies is graphene with an 1 × 1 supercell for 14 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Performance of GAs plotted against GPU hours and across different subsample size k. Averaging over diagonal translations. The results in [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: A G˜ -symmetric configuration of 12 electrons, where G˜ is obtained from P6mm 16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Visualization of the full P6mm-diagonal invariance of ψ (OG) θ versus ψ (PA;G) θ . Same setup and wavefunctions as [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Illustration of different choices of step functions s, the corresponding λϵ’s and their first two derivatives. ϵ = 0.1 above. (iii) Suppose λϵ and d( • , g(Π0)) are p-times continuously differentiable for all g ∈ G, and that ψθ is p-times continuously differentiable a…
Figure 11
Figure 11. Figure 11: Plots of w˜ϵ(x − 1), w˜ϵ(x) and w˜ϵ(x + 1) defined via either s∞ or s2 as the smoothed step function used in λϵ and d, and for different values of ϵ. E.2. An one-electron example of ψ (SC) θ;ϵ For simplicity, we first illustrate ψ (SC) θ;ϵ for a single 1d-electron sys…
Figure 12
Figure 12. Figure 12: Plots of derivatives of w˜ϵ(x) for different values of ϵ. x −1 x−ϵ 0 xϵ x1−ϵ 1 x1+ϵ 2 Π0 [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 14
Figure 14. Figure 14: Plots of d(x, [0, 1)) and its derivatives under different choices of the step function s. F. Proof of Fact 2.1 Fix g ∈ G with its action on x ∈ R 3 represented by g(x) = Ax + b. Given that (ψ, E) solves (1), we seek to show that under the stated conditions, ψg(x) := ψ…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 40 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    J., Bambrick, J., et al

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A. J., Bambrick, J., et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, 630 0 (8016): 0 493--500, 2024

    2024

  3. [3]

    Adams, R. P. and Orbanz, P. Representing and learning functions invariant under crystallographic groups. arXiv preprint arXiv:2306.05261, 2023

    work page Pith review arXiv 2023

  4. [4]

    and Mermin, N

    Ashcroft, N. and Mermin, N. Solid State Physics. Cengage Learning, 1976

    work page 1976

  5. [5]

    The effects of regularization and data augmentation are class dependent

    Balestriero, R., Bottou, L., and LeCun, Y. The effects of regularization and data augmentation are class dependent. In Conference on Neural Information Processing Systems, 2022

    work page 2022

  6. [6]

    P., Simm, G., Ortner, C., and Cs \'a nyi, G

    Batatia, I., Kovacs, D. P., Simm, G., Ortner, C., and Cs \'a nyi, G. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Advances in Neural Information Processing Systems, 35: 0 11423--11436, 2022

    work page 2022

  7. [7]

    P., Kornbluth, M., Molinari, N., Smidt, T

    Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J. P., Kornbluth, M., Molinari, N., Smidt, T. E., and Kozinsky, B. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13 0 (1): 0 2453, 2022

    work page 2022

  8. [8]

    U ber die quantenmechanik der elektronen in kristallgittern. Zeitschrift f \

    Bloch, F. \"U ber die quantenmechanik der elektronen in kristallgittern. Zeitschrift f \"u r physik , 52 0 (7): 0 555--600, 1929

    work page 1929

Show all 53 references
  1. [9]

    P., Hahn, T., Wondratschek, H., M \"u ller, U., Shmueli, U., Prince, E., Authier, A., Kopsk \`y , V., Litvin, D., Arnold, E., et al

    Brock, C. P., Hahn, T., Wondratschek, H., M \"u ller, U., Shmueli, U., Prince, E., Authier, A., Kopsk \`y , V., Litvin, D., Arnold, E., et al. International tables for crystallography volume A: Space-group symmetry. Wiley Online Library, 2016

    2016

  2. [10]

    and Troyer, M

    Carleo, G. and Troyer, M. Solving the quantum many-body problem with artificial neural networks. Science, 355 0 (6325): 0 602--606, 2017

    2017

  3. [11]

    S., and Foulkes, W

    Cassella, G., Sutterud, H., Azadi, S., Drummond, N., Pfau, D., Spencer, J. S., and Foulkes, W. M. C. Discovering quantum phase transitions with fermionic neural networks. Physical Review Letters, 130 0 (3): 0 036401, 2023

    2023

  4. [12]

    H., Goldstein, L., and Shao, Q.-M

    Chen, L. H., Goldstein, L., and Shao, Q.-M. Normal approximation by Stein's method , volume 2. Springer Science & Business Media, Berlin, 2011

    2011

  5. [13]

    Chen, S., Dobriban, E., and Lee, J. H. A group-theoretic framework for data augmentation. Journal of Machine Learning Research, 21 0 (245): 0 1--71, 2020

    2020

  6. [14]

    Central limit theorems and bootstrap in high dimensions

    Chernozhukov, V., Chetverikov, D., and Kato, K. Central limit theorems and bootstrap in high dimensions. Annals of Probability, 45 0 (4): 0 2309--2352, 2017

    2017

  7. [15]

    S E (3) equivariant graph neural networks with complete local frames

    Du, W., Zhang, H., Du, Y., Meng, Q., Chen, W., Zheng, N., Shao, B., and Liu, T.-Y. S E (3) equivariant graph neural networks with complete local frames. In International Conference on Machine Learning, pp.\ 5583--5608. PMLR, 2022

    2022

  8. [16]

    Improved contrastive divergence training of energy based models

    Du, Y., Li, S., Tenenbaum, J., and Mordatch, I. Improved contrastive divergence training of energy based models. arXiv preprint arXiv:2012.01316, 2020

    2012 arXiv

  9. [17]

    A., Schmidt, V., Hern \'a ndez-Garc a, A., Miret, S., Malliaros, F

    Duval, A. A., Schmidt, V., Hern \'a ndez-Garc a, A., Miret, S., Malliaros, F. D., Bengio, Y., and Rolnick, D. Faenet: Frame averaging equivariant gnn for materials modeling. In International Conference on Machine Learning, pp.\ 9013--9033. PMLR, 2023

    2023

  10. [18]

    Dym, N., Lawrence, H., and Siegel, J. W. Equivariant frames and the impossibility of continuous canonicalization. arXiv preprint arXiv:2402.16077, 2024

    2024 arXiv

  11. [19]

    Fundamental high-pressure calibration from all-electron quantum monte carlo calculations

    Esler, K., Cohen, R., Militzer, B., Kim, J., Needs, R., and Towler, M. Fundamental high-pressure calibration from all-electron quantum monte carlo calculations. Physical Review Letters, 104 0 (18): 0 185702, 2010

    2010

  12. [20]

    and G \"u nnemann, S

    Gao, N. and G \"u nnemann, S. Ab-initio potential energy surfaces by pairing gnns with neural wave functions. arXiv preprint arXiv:2110.05064, 2021

    2021 arXiv

  13. [21]

    Gerard, L., Scherbela, M., Marquetand, P., and Grohs, P. Gold-standard solutions to the schr \"o dinger equation using deep learning: How much physics do we need? Advances in Neural Information Processing Systems, 35: 0 10282--10294, 2022

    2022

  14. [22]

    Solving many-electron schr \"o dinger equation using deep neural networks

    Han, J., Zhang, L., et al. Solving many-electron schr \"o dinger equation using deep neural networks. Journal of Computational Physics, 399: 0 108929, 2019

    2019

  15. [23]

    a tzle, Z., and No \'e , F. Deep-neural-network solution of the electronic schr \

    Hermann, J., Sch \"a tzle, Z., and No \'e , F. Deep-neural-network solution of the electronic schr \"o dinger equation. Nature Chemistry, 12 0 (10): 0 891--897, 2020

    2020

  16. [24]

    Hinton, G. E. Training products of experts by minimizing contrastive divergence. Neural computation, 14 0 (8): 0 1771--1800, 2002

    2002

  17. [25]

    H., Orbanz, P., and Austern, M

    Huang, K. H., Orbanz, P., and Austern, M. Gaussian and non-gaussian universality of data augmentation. arXiv preprint arXiv:2202.09134, 1, 2022

    2022

  18. [26]

    H., Liu, X., Duncan, A., and Gandy, A

    Huang, K. H., Liu, X., Duncan, A., and Gandy, A. A high-dimensional convergence theorem for u-statistics with applications to kernel-based testing. In The Thirty Sixth Annual Conference on Learning Theory, pp.\ 3827--3918. PMLR, 2023

    2023

  19. [27]

    P., Hautier, G., Chen, W., Richards, W

    Jain, A., Ong, S. P., Hautier, G., Chen, W., Richards, W. D., Dacek, S., Cholia, S., Gunter, D., Skinner, D., Ceder, G., et al. Commentary: The materials project: A materials genome approach to accelerating materials innovation. APL materials, 1 0 (1), 2013

    2013

  20. [28]

    K., Zhang, Y., Bengio, Y., and Ravanbakhsh, S

    Kaba, S.-O., Mondal, A. K., Zhang, Y., Bengio, Y., and Ravanbakhsh, S. Equivariance with learned canonicalization functions. In International Conference on Machine Learning, pp.\ 15546--15566. PMLR, 2023

    2023

  21. [29]

    R., Needs, R., and Rajagopal, G

    Kent, P. R., Needs, R., and Rajagopal, G. Monte carlo energy and variance-minimization techniques for optimizing many-body wave functions. Physical Review B, 59 0 (19): 0 12344, 1999

    1999

  22. [30]

    and McEuen, P

    Kittel, C. and McEuen, P. Introduction to solid state physics. John Wiley & Sons, 2018

    2018

  23. [31]

    L., Romano, J

    Lehmann, E. L., Romano, J. P., and Casella, G. Testing statistical hypotheses, volume 3. Springer, 1986

    1986

  24. [32]

    A computational framework for neural network-based variational monte carlo with forward laplacian

    Li, R., Ye, H., Jiang, D., Wen, X., Wang, C., Li, Z., Li, X., He, D., Chen, J., Ren, W., et al. A computational framework for neural network-based variational monte carlo with forward laplacian. Nature Machine Intelligence, 6 0 (2): 0 209--219, 2024

    2024

  25. [33]

    Ab initio calculation of real solids via neural network ansatz

    Li, X., Li, Z., and Chen, J. Ab initio calculation of real solids via neural network ansatz. Nature Communications, 13 0 (1): 0 7895, 2022

    2022

  26. [34]

    Explicitly antisymmetrized neural network layers for variational monte carlo simulation

    Lin, J., Goldshlager, G., and Lin, L. Explicitly antisymmetrized neural network layers for variational monte carlo simulation. Journal of Computational Physics, 474: 0 111765, 2023

    2023

  27. [35]

    M., and Clark, B

    Luo, D., Chen, Z., Hu, K., Zhao, Z., Hur, V. M., and Clark, B. K. Gauge-invariant and anyonic-symmetric autoregressive neural network for quantum lattice models. Physical Review Research, 5 0 (1): 0 013216, 2023

    2023

  28. [36]

    On the benefits of invariance in neural networks

    Lyle, C., van der Wilk, M., Kwiatkowska, M., Gal, Y., and Bloem-Reddy, B. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020

    2005 arXiv

  29. [37]

    and Sharma, S

    Mahajan, A. and Sharma, S. Symmetry-projected jastrow mean-field wave function in variational monte carlo. The Journal of Physical Chemistry A, 123 0 (17): 0 3911--3921, 2019

    2019

  30. [38]

    and Grosse, R

    Martens, J. and Grosse, R. Optimizing neural networks with kronecker-factored approximate curvature. In International conference on machine learning, pp.\ 2408--2417. PMLR, 2015

    2015

  31. [39]

    Integral probability metrics and their generating classes of functions

    M \"u ller, A. Integral probability metrics and their generating classes of functions. Advances in applied probability, 29 0 (2): 0 429--443, 1997

    1997

  32. [40]

    P., Burke, K., and Ernzerhof, M

    Perdew, J. P., Burke, K., and Ernzerhof, M. Generalized gradient approximation made simple. Physical review letters, 77 0 (18): 0 3865, 1996

    1996

  33. [41]

    S., Matthews, A

    Pfau, D., Spencer, J. S., Matthews, A. G., and Foulkes, W. M. C. Ab initio solution of the many-electron schr \"o dinger equation with deep neural networks. Physical review research, 2 0 (3): 0 033429, 2020

    2020

  34. [42]

    and Krishnapriyan, A

    Qu, E. and Krishnapriyan, A. The importance of being scalable: Improving the speed and accuracy of neural network interatomic potentials across chemical domains. Advances in Neural Information Processing Systems, 37: 0 139030--139053, 2024

    2024

  35. [43]

    Variational and diffusion quantum monte carlo calculations at nonzero wave vectors: Theory and application to diamond-structure germanium

    Rajagopal, G., Needs, R., James, A., Kenny, S., and Foulkes, W. Variational and diffusion quantum monte carlo calculations at nonzero wave vectors: Theory and application to diamond-structure germanium. Physical Review B, 51 0 (16): 0 10591, 1995

    1995

  36. [44]

    Foundations of hyperbolic manifolds

    Ratcliffe, J. Foundations of hyperbolic manifolds. Graduate Texts in Mathematics/Springer-Verlag, 149, 1994

    1994

  37. [45]

    Optimization of the structure of the moment bounds for accuracy of normal approximation for the distributions of sums of independent random variables

    Shevtsova, I. Optimization of the structure of the moment bounds for accuracy of normal approximation for the distributions of sums of independent random variables. PhD thesis, Moscow State University, 2013

    2013

  38. [46]

    Group theory and quantum mechanics

    Tinkham, M. Group theory and quantum mechanics. Courier Corporation, 2003

    2003

  39. [47]

    S., and Pfau, D

    von Glehn, I., Spencer, J. S., and Pfau, D. A self-attention ansatz for ab-initio quantum chemistry. In The Eleventh International Conference on Learning Representations, 2023

    2023

  40. [48]

    Jastrow correlation factor for periodic systems

    Whitehead, T., Michael, M., and Conduit, G. Jastrow correlation factor for periodic systems. Physical Review B, 94 0 (3): 0 035157, 2016

    2016

  41. [49]

    T., Yao, Y., Li, J., Chen, L., Shi, H., Motta, M., Niu, C., Ray, U., Guo, S., Anderson, R

    Williams, K. T., Yao, Y., Li, J., Chen, L., Shi, H., Motta, M., Niu, C., Ray, U., Guo, S., Anderson, R. J., et al. Direct comparison of many-body methods for realistic electronic hamiltonians. Physical Review X, 10 0 (1): 0 011041, 2020

    2020

  42. [50]

    Pseudopotential variational quantum monte carlo approach to bcc lithium

    Yao, G., Xu, J., and Wang, X. Pseudopotential variational quantum monte carlo approach to bcc lithium. Physical Review B, 54 0 (12): 0 8393, 1996

    1996

  43. [51]

    Deep density: circumventing the kohn-sham equations via symmetry preserving neural networks

    Zepeda-N \'u \ n ez, L., Chen, Y., Zhang, J., Jia, W., Zhang, L., and Lin, L. Deep density: circumventing the kohn-sham equations via symmetry preserving neural networks. Journal of Computational Physics, 443: 0 110523, 2021

    2021

  44. [52]

    o dingernet: A universal neural network solver for the schr \

    Zhang, Y., Jiang, B., and Guo, H. Schro \"o dingernet: A universal neural network solver for the schr \"o dinger equation. Journal of Chemical Theory and Computation, 21 0 (2): 0 670--677, 2025

    2025

  45. [53]

    and Dovesi, R

    Zicovich-Wilson, C. and Dovesi, R. On the use of symmetry-adapted crystalline orbitals in scf-lcao periodic calculations. ii. implementation of the self-consistent-field scheme and examples. International journal of quantum chemistry, 67 0 (5): 0 311--320, 1998

    1998

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.