Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Machine Learning Gravity Compactifications on Negatively Curved Manifolds

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A neural-network ansatz deforms a piecewise Einstein metric into a smooth Einstein metric on a Dehn-filled hyperbolic three-manifold, cutting error on every equation by orders of magnitude.

desk verdict A careful, honest proof-of-concept that ML can solve Einstein equations on a Dehn-filled cusped 3-manifold, with a validation that is weaker than the text suggests. read the letter →

arxiv 2501.00093 v1 pith:PYPQNNA3 submitted 2024-12-30 hep-th cs.LGgr-qc

G. Bruno De Luca This is my paper · ORCID
classification hep-thcs.LGgr-qc
keywords machinelearningPDEsgravitycompactificationsEinsteinmetricsDehnfillinghyperbolicmanifoldsneuralnetworksM-theorydeSittervacuanegativelycurved
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 argues that direct machine-learning solution of the Einstein equations is a viable route to constructing gravity compactifications without relying on supersymmetry or special analytic structures. As a proof of concept, it takes a cusped hyperbolic three-manifold, fills one cusp by the Dehn-filling construction, and uses six neural networks, pre-trained on the resulting piecewise Einstein metric, to deform it into a smooth solution of the Einstein equation while enforcing matching of induced metric and extrinsic curvature on all identified facets. The trained metric reduces the volume-averaged Einstein-equation error to $(9.7\pm0.5)\times10^{-3}$ and the boundary-condition errors to $10^{-4}$ to $10^{-5}$, which the author reads as evidence of convergence to the true solution. Since the filling construction and the neural-network ansatz work for dimensions $3\le n\le8$, the same method is proposed as a path toward new negatively curved Einstein metrics and toward numerical construction of four-dimensional de Sitter compactifications of M-theory.

What carries the argument

The load-bearing machinery is the Dehn-filling construction paired with a neural-network metric parametrization. The filling cuts a cusp of a hyperbolic manifold at a chosen gluing radius, caps it with a Euclidean AdS black hole metric, and produces a piecewise Einstein metric that is continuous but not smooth; the Dehn-filling existence theorem guarantees a nearby smooth Einstein metric. The networks implement that deformation: each of the six independent metric components is a small residual network (width 10, depth 2, 1986 trainable parameters), with prefactors such as $z^{-2}(r_{\rm join}-z)^{-1}$ and $(r_{\rm join}-z)z^{-2}$ removing the coordinate singularities, so the networks learn only the smooth part. Pre-training matches the piecewise metric in value and in first and second $z$-derivatives, and the training loss then sums the bulk Einstein equation with the 14 identified-facet conditions, using hyperbolic-volume sampling with resampling of high-error points. This turns the analytic existence proof into a concrete numerical construction.

What would settle it

Recompute the volume-averaged relative error using the learned metric's own volume form on a dense independent grid, especially in regions where the fiducial piecewise volume density is small; if the self-weighted error is far above $10^{-3}$ while the fiducial-weighted error remains small, the claimed convergence is an artifact of the weighting.

Watch

Extended reading notes

Core claim

In the author's own terms, the discovery is that a class of Einstein metrics whose existence is guaranteed by the Dehn-filling theorem can actually be constructed numerically by a machine-learning deformation of the explicit piecewise metric. On the manifold $M_3$, built from the right-angled polytope $P_3$ via a three-coloring and facet identifications, the paper solves $R_{mn}+2\delta^{-1}g_{mn}=0$ on the bulk of $P_3^8$ while imposing the 14 pairs of continuity and extrinsic-curvature matching conditions. The metric ansatz factors out the coordinate singularities so that the six fully connected networks only need to learn the smooth part of $g_{mn}$. After pre-training on the piecewise metric (3.19) and then training on the full loss, the reported errors are about $10^{-2}$ for the Einstein equation and $10^{-4}$ to $10^{-5}$ for the junction conditions, orders of magnitude below the two manual smooth interpolations used as baselines. Lacking an independent ground truth, the paper treats these error reductions as evidence of convergence.

Load-bearing premise

The convergence claim assumes that using the fixed piecewise metric to define the volume weight in the error averages does not hide large Einstein-equation violations in regions where that fiducial metric assigns little volume, since no independent ground-truth solution is available for comparison.

Editorial extensions

If this is right

  • Because the neural-network ansatz never uses the three-dimensional fact that an Einstein metric is locally hyperbolic, the same loss construction transfers directly to dimensions 4 through 8, where the analogous filled Einstein metrics are known to exist but have not been explicitly constructed.
  • Starting from a patchwise approximate solution is presented as a general strategy: pre-training the networks to approximate it gives an initialization from which the full PDE training converges, so approximate solutions from standard methods can be upgraded to numerical solutions.
  • The same machinery can be pointed at the full compactification system with warping and matter fields, because the loss function simply sums the bulk equations, the matter equations, and the interface conditions, removing the need to invoke supersymmetry to solve the equations of motion.
  • An explicit numerical Einstein metric on a filled manifold would make the internal geometry of the proposed dS4 M-theory compactifications computable, yielding the cycle volumes that control Casimir energies and hence the effective potential.
  • The training procedure also provides a benchmark for optimizers on high-dimensional PDE losses, since different successful runs and optimizers reach comparable average relative errors on this three-dimensional testbed.

Reading between the lines

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

  • If the method scales as the paper suggests, the practical bottleneck shifts from solving the PDEs to producing a reliable approximate starting metric and an error measure that networks cannot shrink by concentrating volume; a natural safeguard would be to weight errors by the learned metric's own volume and monitor the ratio of volume elements.
  • The same pre-train-then-solve recipe could be applied to other patchwise-defined geometric PDE problems, such as numerical Kähler-Einstein metrics or brane backreaction, whenever an approximate gluing of local solutions is available.
  • A testable near-term extension is to run the same code on a filling with more than one cusp and check that per-cusp errors stay small independently; if they do not, the interface conditions may need larger weight or a gauge-fixing term.
  • Because the paper reports that several optimizers reach comparable errors in this three-dimensional problem, higher-dimensional tests will be needed to determine whether optimizer choice matters on rougher loss landscapes; that question is left open by this work.
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

3 major / 5 minor

Summary. The paper proposes a physics-informed machine-learning pipeline for solving Einstein-like PDEs on patchwise-defined compactification manifolds. After reviewing the warped compactification equations and the Anderson-Dehn filling construction, it specializes to a three-dimensional cusped hyperbolic manifold M3, constructs an explicit piecewise Einstein metric (3.19), parameterizes a smooth deformation by six fully connected neural networks (3.25), pre-trains the network to the piecewise metric, and then minimizes the Einstein residual (3.23) together with continuity and extrinsic-curvature matching conditions (3.24). The main reported result is Table 1: a trained Einstein-equation average relative error of (9.7 ± 0.5) × 10^-3 and boundary-condition average relative errors of order 10^-4 to 10^-5, which the paper interprets as evidence of convergence to the true smooth Einstein metric. The code is released under a GPL license.

Significance. If the central claim is established, this is a useful proof of concept that machine learning can directly solve nonlinear geometric PDEs on manifolds defined by facet identifications, with potential applicability to higher-dimensional negatively curved Einstein metrics and eventually to M-theory de Sitter compactifications. The paper has several genuine strengths: the boundary treatment is explicit, with the 14 identified facet pairs and their Jacobians given in Appendix C; the pre-training strategy provides a practical way to feed in an approximate solution; the metric ansatz in (3.25) builds in the coordinate singularities explicitly; and the code is publicly available. The main caveat is that the evidence for convergence is internal: the validation uses weighted residuals of the same equations being trained, with no independent numerical or geometric ground truth. For this reason the language 'convergence to the true solution' is stronger than what the present data establish.

major comments (3)
  1. [Sec. 3.3.1, after Eq. (3.24)] The ARE used to support the central convergence claim is weighted by the volume form of the fixed piecewise metric, not by the learned metric or by a uniform measure; footnote 9 explicitly acknowledges this choice. A small value of (3.28) therefore certifies small residuals only in regions that are heavily weighted by the fiducial volume form, and if the learned and fiducial metrics differ in low-fiducial-volume regions, large violations of (3.23) and (3.24) could remain invisible to the reported number. The paper does not provide the maximum local relative error, an unweighted average, or a comparison with an independent numerical solution, and Sec. 3.3.2 explicitly states that no ground truth is available. I request at least one additional validation, such as a pointwise error map, a region-by-region decomposition of the ARE, or a comparison of geometric invariants (volume, closed geodesic lengths) with the known filling. I note that I do not find the specific claim that the fiducial volume form diverges as 1/sqrt(rjoin - z) to be correct: the determinant of (3.19) in the filled region is finite at z = rjoin. The weighting concern nevertheless stands, because the error measure and the training loss are the same residuals and the validation measure is not independent of the training procedure.
  2. [Sec. 3.3.1, after Eq. (3.24)] The manuscript explicitly states that no gauge constraint was imposed: 'For this simple three-dimensional problem, we did not find necessary to fix the gauge constraint ξ = 0.' Since the Einstein equation (3.23) is invariant under diffeomorphisms, the optimization landscape has flat directions along gauge transformations, and a small final residual does not by itself identify a unique coordinate representative of the metric. This is not fatal, because the pre-training to the piecewise metric (3.19) and the boundary conditions (3.24) may select the intended representative, but that selection is not demonstrated. I recommend either imposing the DeTurck gauge discussed in Sec. 2.2, or reporting the magnitude of the gauge vector ξ (or an equivalent check) for the trained solution to show that gauge drift is controlled.
  3. [Sec. 3.3.2, Table 1] The statement that the trained network 'is able to reduce the error simultaneously on the Einstein equations and all the boundary conditions by many orders of magnitude' overstates the Einstein-equation improvement. Table 1 shows the Einstein ARE decreasing from 0.121-0.145 for the manual interpolations to (9.7 ± 0.5) × 10^-3, a factor of roughly 15, while the boundary-condition AREs improve by several orders of magnitude. The wording should be adjusted to distinguish the two improvements, or the claim should be restricted to the boundary conditions.
minor comments (5)
  1. [Sec. 3.3.1, Eq. (3.23)] Please clarify whether the constant δ is fixed to 1 or is a trainable or tunable parameter; both the equation (3.23) and the pre-training target δ × (3.19) depend on its value.
  2. [Fig. 4] The axis label 'R/(6 1)+1' appears to contain a typesetting error; it should presumably be R/(6δ^{-1}) + 1 or an explicitly defined dimensionless ratio.
  3. [Sec. 3.2.2, Eqs. (3.17) and (3.19)] The text says that for z ≥ 1 the hyperbolic metric reads (3.17), but the piecewise metric (3.19) uses the same hyperbolic expression for 0 ≤ z ≤ 1 and the filled expression for 1 ≤ z ≤ rjoin; a sentence explaining the relation between these ranges would remove an apparent inconsistency.
  4. [Sec. 4] There are a few typos, including 'pertubative' and 'dimensionalty', which should be corrected.
  5. [References] Reference [73] is cited as 'Work in progress' with an empty title; if it is needed, please provide a placeholder title or remove the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported errors are the training objective evaluated on fresh samples, not a fitted quantity masquerading as a prediction.

full rationale

The central derivation is not circular. The neural-network metric is obtained by minimizing the Einstein-residual and boundary-condition loss (2.11), with the Einstein equation E in (3.23) and boundary conditions B in (3.24), starting from a pre-trained smooth approximation to the piecewise metric (3.19). The final metric is not fitted to the validation ARE; the AREs in Table 1 are Monte Carlo evaluations of the same relative-error norm (3.27)-(3.28) on newly sampled points. Minimizing a residual and then reporting that residual is a self-consistency check rather than an independent prediction, and because the validation volume form in footnote 9 (the fixed piecewise metric) differs from the hyperbolic volume form used for training sampling in the filled patch, the validation is not identical to the training loss by construction. The piecewise metric serves only as initialization, not as the target of the reported improvement. Self-citations to [39], [45], [46], and [41] supply physical motivation and an optimizer, but the method is also tested with Adam, and the existence of a nearby smooth Einstein metric is an external theorem from [48,49]; no load-bearing assumption is justified solely by the author's own prior work. The footnote-9 weighting caveat is a genuine limitation for pointwise certification, but it is a numerical-validation weakness, not a circular reduction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on established mathematical existence theorems (Anderson-Dehn filling), the explicit polytope construction from [50], and the empirical performance of neural-network optimization. The only hand-chosen numbers are the conformal factor scale and the network/training hyperparameters.

free parameters (2)
  • Conformal factor delta = Not reported (pre-training to piecewise metric fixes the scale)
    The Einstein equation is written as Rmn + 2/delta gmn = 0 with delta an arbitrary constant conformal factor; the overall scale of the solution is fixed by the pre-training target (3.19) with ell = 1, but the paper does not state delta explicitly.
  • Network and training hyperparameters = H=10, D=2, pre-training threshold 3e-3, gamma_E=gamma_B=1, 60% error-retention
    Chosen by hand; no sensitivity analysis is reported, so the demonstrated feasibility could depend on these choices.
assumptions (4)
  • domain assumption Smooth Einstein metrics obtained by Anderson-Dehn filling of cusped hyperbolic manifolds exist (Anderson 2006; Bamler 2012).
    The paper relies on this theorem to know a smooth solution exists near the piecewise metric (3.19); it is not proved in this work.
  • domain assumption The right-angled polytope P3 and the 3-coloring construction from Italiano-Martelli-Migliorini produce the finite-volume hyperbolic manifold M3 with identified facets as described.
    Used in Sec. 3.2 to define the domain and boundary identifications; taken from [50].
  • domain assumption The neural-network ansatz (3.25) can represent the true smooth Einstein metric within the required accuracy, and the optimizers can reach a loss minimum with nearly zero residual.
    The paper notes ML methods lack convergence guarantees (Sec. 1); the claim rests on empirical evidence from a limited number of runs.
  • domain assumption For this 3d problem, gauge fixing (xi = 0) is unnecessary and the boundary conditions (3.20), (3.22) suffice to make the problem well-posed.
    Stated in Sec. 3.3.1 without proof; while plausible for a three-dimensional deformation, it is a modeling assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning Gravity Compactifications on Negatively Curved Manifolds." pith.science (2026). https://pith.science/paper/PYPQNNA3

@misc{pith2026250100093,
  author       = {Pith},
  title        = {Pith review of: Machine Learning Gravity Compactifications on Negatively Curved Manifolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PYPQNNA3}},
  note         = {Machine review of arXiv:2501.00093}
}
read the original abstract

Constructing the landscape of vacua of higher-dimensional theories of gravity by directly solving the low-energy (semi-)classical equations of motion is notoriously difficult. In this work, we investigate the feasibility of Machine Learning techniques as tools for solving the equations of motion for general warped gravity compactifications. As a proof-of-concept we use Neural Networks to solve the Einstein PDEs on non-trivial three manifolds obtained by filling one or more cusps of hyperbolic manifolds. While in three dimensions an Einstein metric is also locally hyperbolic, the generality and scalability of Machine Learning methods, the availability of explicit families of hyperbolic manifolds in higher dimensions, and the universality of the filling procedure strongly suggest that the methods and code developed in this work can be of broader applicability. Specifically, they can be used to tackle both the geometric problem of numerically constructing novel higher-dimensional negatively curved Einstein metrics, as well as the physical problem of constructing four-dimensional de Sitter compactifications of M-theory on the same manifolds.

Figures

Figures reproduced from arXiv: 2501.00093 by the authors.

Figure 1
Figure 1. An embedding of P3 in the UHS model (3.13) with z the vertical coordinate. This polytope is composed of 6 totally geodesics facets: four vertical walls and two hemispheres centered at z = 0. In this picture the coordinate z has been cut-off at a finite volume, but the polytope extends up to z = ∞, while remaining of finite volume with respect to the metric (3.13). subgroup Γ ′ ≤ Γ whose action on H3 is free. A stand… view at source ↗
Figure 2
Figure 2. A view from the top of the larger polytope [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 3
Figure 3. A typical training run. Left: pre-training loss when training to approximate the piece-wise solution starting from a random initialization, right: training loss (running average) when training on the full loss starting from the pre-trained network. The pre-training is early￾stopped when the pre-training loss reaches a certain threshold, as its role is only to serve as a good initialization for the actual training ru… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Percentage deviation of the Ricci scalar from its true value for a k = 5 filling, evalu￾ated on the slice x1 = x2 = 1 2 , after the pre-training phase (green) and after the training phase of a typical training run (purple), compared to two different manual smooth inter…
Figure 5
Figure 5. Figure 5: Local Relative Error for all the components of the Einstein equation on a typical training run. Shaded area shows 5% error. 1. For each Ri , we generate all the extra reflections generated by the new added facets Ra i , where a denotes an internal face generated by a r…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Black Hole Black Boxes: Numerical Black Hole Metrics via AInstein Neural Networks

    gr-qc 2026-07 conditional novelty 6.0 of 10

    Unsupervised Lorentzian PINNs with embedded S^{2} topology recover maximally extended Schwarzschild and yield candidate Petrov type-I vacuum black-hole metrics with genuinely trapped interiors.

Reference graph

Works this paper leans on

75 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [1]

    De Sitter vacua in string theory,

    S. Kachru, R. Kallosh, A. D. Linde, and S. P. Trivedi, “De Sitter vacua in string theory,” Phys. Rev. D68 (2003) 046005, hep-th/0301240

    arXiv 2003

  2. [2]

    Breaking supersymmetry with pure spinors,

    A. Legramandi and A. Tomasiello, “Breaking supersymmetry with pure spinors,” Journal of High Energy Physics2020 (2020), no. 11, 1–25

    work page 2020

  3. [3]

    Tomasiello,Geometry of String Theory Compactifications

    A. Tomasiello,Geometry of String Theory Compactifications. Cambridge Univer- sity Press, 1, 2022

    work page 2022

  4. [4]

    Polchinski and E

    J. Polchinski and E. Silverstein, Dual Purpose Landscaping Tools: Small Extra Dimensions in AdS/CFT, pp. 365–390. 8, 2009.0908.0756

    arXiv 2009

  5. [5]

    Type IIA moduli stabiliza- tion,

    O. DeWolfe, A. Giryavets, S. Kachru, and W. Taylor, “Type IIA moduli stabiliza- tion,” JHEP 07 (2005) 066, hep-th/0505160

    arXiv 2005

  6. [6]

    Scale- separated AdS4 vacua of IIA orientifolds and M-theory,

    N. Cribiori, D. Junghans, V. Van Hemelryck, T. Van Riet, and T. Wrase, “Scale- separated AdS4 vacua of IIA orientifolds and M-theory,”Phys. Rev. D104 (2021), no. 12, 126014,2107.00019

    arXiv 2021

  7. [7]

    Small cos- mological constants in string theory,

    M. Demirtas, M. Kim, L. McAllister, J. Moritz, and A. Rios-Tascon, “Small cos- mological constants in string theory,”JHEP 12 (2021) 136, 2107.09064

    arXiv 2021

  8. [8]

    Gravity from ther- modynamics: Optimal transport and negative effective dimensions,

    G. B. De Luca, N. De Ponti, A. Mondino, and A. Tomasiello, “Gravity from ther- modynamics: Optimal transport and negative effective dimensions,”SciPost Phys. 15 (2023), no. 2, 039,2212.02511

    arXiv 2023

Show all 75 references
  1. [9]

    Quantum corrections to DGKT and the Weak Gravity Conjecture,

    M. Montero and I. Valenzuela, “Quantum corrections to DGKT and the Weak Gravity Conjecture,” 2412.00189

    arXiv

  2. [10]

    Snowmass White Paper: Cosmology at the Theory Frontier,

    R. Flauger, V. Gorbenko, A. Joyce, L. McAllister, G. Shiu, and E. Silverstein, “Snowmass White Paper: Cosmology at the Theory Frontier,” inSnowmass 2021. 3, 2022. 2203.07629

    2021 arXiv

  3. [11]

    TASI lectures on cosmological observables and string theory,

    E. Silverstein, “TASI lectures on cosmological observables and string theory,” in Theoretical Advanced Study Institute in Elementary Particle Physics: New Fron- tiers in Fields and Strings, pp. 545–606. 2017.1606.03640. 37

    2017

  4. [12]

    Moduli Stabilization in String Theory,

    L. McAllister and F. Quevedo, “Moduli Stabilization in String Theory,” 2310.20559

    arXiv

  5. [13]

    New de Sitter Solutions,

    E. Silverstein, “New de Sitter Solutions,”Plenary Talk at Strings, June 2013

    2013

  6. [14]

    Classical de Sitter Solutions of 10-Dimensional Supergravity,

    C. Córdova, G. B. De Luca, and A. Tomasiello, “Classical de Sitter Solutions of 10-Dimensional Supergravity,” Phys. Rev. Lett. 122 (2019), no. 9, 091601, 1812.04147

    2019 arXiv

  7. [15]

    New de Sitter Solutions in Ten Dimensions and Orientifold Singularities,

    C. Córdova, G. B. De Luca, and A. Tomasiello, “New de Sitter Solutions in Ten Dimensions and Orientifold Singularities,”JHEP 08 (2020) 093, 1911.04498

    2020 arXiv

  8. [16]

    4D de Sitter from String Theory via 6D Supergravity,

    C. P. Burgess, F. Muia, and F. Quevedo, “4D de Sitter from String Theory via 6D Supergravity,” 2408.03852

    arXiv

  9. [17]

    AdS8 solutions in type II super- gravity,

    C. Córdova, G. B. De Luca, and A. Tomasiello, “AdS8 solutions in type II super- gravity,” JHEP 07 (2019) 127, 1811.06987

    2019 arXiv

  10. [18]

    A New approach to static numerical relativity, and its application to Kaluza-Klein black holes,

    M. Headrick, S. Kitchen, and T. Wiseman, “A New approach to static numerical relativity, and its application to Kaluza-Klein black holes,”Class. Quant. Grav.27 (2010) 035002, 0905.1822

    2010 arXiv

  11. [19]

    Numerical Ricci-flat metrics on K3,

    M. Headrick and T. Wiseman, “Numerical Ricci-flat metrics on K3,”Class. Quant. Grav.22 (2005) 4931–4960, hep-th/0506129

    2005 arXiv

  12. [20]

    Numerical Calabi-Yau metrics,

    M. R. Douglas, R. L. Karp, S. Lukic, and R. Reinbacher, “Numerical Calabi-Yau metrics,” J. Math. Phys.49 (2008) 032302, hep-th/0612075

    2008 arXiv

  13. [21]

    Energy functionals for Calabi-Yau metrics,

    M. Headrick and A. Nassar, “Energy functionals for Calabi-Yau metrics,” Adv. Theor. Math. Phys.17 (2013), no. 5, 867–902,0908.2635

    2013 arXiv

  14. [22]

    Numerical Kahler-Einstein metric on the third del Pezzo,

    C. Doran, M. Headrick, C. P. Herzog, J. Kantor, and T. Wiseman, “Numerical Kahler-Einstein metric on the third del Pezzo,”Commun. Math. Phys.282 (2008) 357–393, hep-th/0703057

    2008 arXiv

  15. [23]

    Physics-informed machine learning,

    G. Karniadakis, Y. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,”Nature Reviews Physics(05, 2021) 1–19

    2021

  16. [24]

    Artificial neural networks for solving ordinary and partial differential equations,

    I. E. Lagaris, A. Likas, and D. I. Fotiadis, “Artificial neural networks for solving ordinary and partial differential equations,”IEEE transactions on neural networks 9 (1998), no. 5, 987–1000

    1998

  17. [25]

    Bootstrapping string models with entanglement minimization and Machine-Learning,

    F. Bhat, D. Chowdhury, A. P. Saha, and A. Sinha, “Bootstrapping string models with entanglement minimization and Machine-Learning,”2409.18259. 38

    arXiv

  18. [26]

    Machine Learning Toric Duality in Brane Tilings,

    P. Capuozzo, T. S. Gherardini, and B. Suzzoni, “Machine Learning Toric Duality in Brane Tilings,”2409.15251

    arXiv

  19. [27]

    The S-matrix bootstrap with neural optimizers I: zero double discontinuity,

    M. A. Gumus, D. Leflot, P. Tourkine, and A. Zhiboedov, “The S-matrix bootstrap with neural optimizers I: zero double discontinuity,”2412.09610

    arXiv

  20. [28]

    Rigor with machine learning from field theory to the Poincaré conjecture,

    S. Gukov, J. Halverson, and F. Ruehle, “Rigor with machine learning from field theory to the Poincaré conjecture,”Nature Rev. Phys.6 (2024), no. 5, 310–319, 2402.13321

    2024 arXiv

  21. [29]

    KAN: Kolmogorov-Arnold Networks,

    Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljačić, T. Y. Hou, and M. Tegmark, “KAN: Kolmogorov-Arnold Networks,”2404.19756

    arXiv

  22. [30]

    Dgm: A deep learning algorithm for solving partial differential equations,

    J. Sirignano and K. Spiliopoulos, “Dgm: A deep learning algorithm for solving partial differential equations,”Journal of computational physics375 (2018) 1339– 1364

    2018

  23. [31]

    Machine Learning Calabi–Yau Metrics,

    A. Ashmore, Y.-H. He, and B. A. Ovrut, “Machine Learning Calabi–Yau Metrics,” Fortsch. Phys.68 (2020), no. 9, 2000068,1910.08605

    2020 arXiv

  24. [32]

    Numerical Calabi-Yau metrics from holomorphic networks,

    M. R. Douglas, S. Lakshminarasimhan, and Y. Qi, “Numerical Calabi-Yau metrics from holomorphic networks,”2012.04797

    2012 arXiv

  25. [33]

    Neural network approximations for Calabi-Yau metrics,

    V. Jejjala, D. K. Mayorga Pena, and C. Mishra, “Neural network approximations for Calabi-Yau metrics,”JHEP 08 (2022) 105, 2012.15821

    2022 arXiv

  26. [34]

    Learning Size and Shape of Calabi-Yau Spaces,

    M. Larfors, A. Lukas, F. Ruehle, and R. Schneider, “Learning Size and Shape of Calabi-Yau Spaces,” Machine Learning and the Physical Sciences, Workshop at 35th NeurIPS(2021) 2111.01436

    2021 arXiv

  27. [35]

    Numericalmetricsforcomplete intersection and Kreuzer–Skarke Calabi–Yau manifolds,

    M.Larfors, A.Lukas, F.Ruehle, andR.Schneider, “Numericalmetricsforcomplete intersection and Kreuzer–Skarke Calabi–Yau manifolds,”Mach. Learn. Sci. Tech. 3 (2022), no. 3, 035014,2205.13408

    2022 arXiv

  28. [36]

    CYJAX: A package for Calabi-Yau metrics with JAX,

    M. Gerdes and S. Krippendorf, “CYJAX: A package for Calabi-Yau metrics with JAX,” Mach. Learn. Sci. Tech.4 (2023), no. 2, 025031,2211.12520

    2023 arXiv

  29. [37]

    Not So Flat Metrics,

    C. S. Fraser-Taliente, T. R. Harvey, and M. Kim, “Not So Flat Metrics,” 2411.00962

    arXiv

  30. [38]

    Lectures on Numerical and Machine Learning Methods for Approximating Ricci-flat Calabi-Yau Metrics,

    L. B. Anderson, J. Gray, and M. Larfors, “Lectures on Numerical and Machine Learning Methods for Approximating Ricci-flat Calabi-Yau Metrics,” 12, 2023. 2312.17125. 39

    2023 arXiv

  31. [39]

    Hyperbolic compactification of M-theory and de Sitter quantum gravity,

    G. B. De Luca, E. Silverstein, and G. Torroba, “Hyperbolic compactification of M-theory and de Sitter quantum gravity,”SciPost Phys. 12 (2022), no. 3, 083, 2104.13380

    2022 arXiv

  32. [40]

    Adam: A method for stochastic optimization,

    D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  33. [41]

    Optimizers for Stabi- lizing Neural Simulation-based Inference,

    G. B. De Luca, B. Nachman, E. Silverstein, and H. Zheng, “Optimizers for Stabi- lizing Neural Simulation-based Inference,”To Appear

  34. [42]

    Microcanonical Hamilto- nian Monte Carlo,

    J. Robnik, G. B. De Luca, E. Silverstein, and U. Seljak, “Microcanonical Hamilto- nian Monte Carlo,”2212.08549

    arXiv

  35. [43]

    Capse.jl: efficient and auto- differentiable CMB power spectra emulation,

    M. Bonici, F. Bianchini, and J. Ruiz-Zapatero, “Capse.jl: efficient and auto- differentiable CMB power spectra emulation,”2307.14339

    arXiv

  36. [44]

    Microcanonical Langevin Monte Carlo,

    J. Robnik and U. Seljak, “Microcanonical Langevin Monte Carlo,”2303.18221

    arXiv

  37. [45]

    Born-Infeld (BI) for AI: Energy-Conserving Descent (ECD) for Optimization,

    G. B. De Luca and E. Silverstein, “Born-Infeld (BI) for AI: Energy-Conserving Descent (ECD) for Optimization,”PMLR 162 (2022) 4918, 2201.11137

    2022 arXiv

  38. [46]

    ImprovingEnergyConservingDescent for Machine Learning: Theory and Practice,

    G.B.DeLuca, A.Gatti, andE.Silverstein, “ImprovingEnergyConservingDescent for Machine Learning: Theory and Practice,”2306.00352

    arXiv

  39. [47]

    The ademamix optimizer: Better, faster, older,

    M. Pagliardini, P. Ablin, and D. Grangier, “The ademamix optimizer: Better, faster, older,” arXiv preprint arXiv:2409.03137(2024)

    2024 arXiv

  40. [48]

    Dehn Filling and Einstein Metrics in Higher Dimensions,

    M. T. Anderson, “Dehn Filling and Einstein Metrics in Higher Dimensions,”Jour- nal of Differential Geometry73 (2006), no. 2, 219 – 261

    2006

  41. [49]

    Constructionofeinsteinmetricsbygeneralizeddehnfilling,

    R.H.Bamler, “Constructionofeinsteinmetricsbygeneralizeddehnfilling,” Journal of the European Mathematical Society14 (2012), no. 3, 887–909

    2012

  42. [50]

    Hyperbolic manifolds that fibre alge- braically up to dimension 8,

    G. Italiano, B. Martelli, and M. Migliorini, “Hyperbolic manifolds that fibre alge- braically up to dimension 8,”Journal of the Institute of Mathematics of Jussieu 23 (2024), no. 2, 609–646

    2024

  43. [51]

    Gnu general public license, http://www.gnu.org/licenses/gpl.html

    “Gnu general public license, http://www.gnu.org/licenses/gpl.html.”

  44. [52]

    Leaps and bounds towards scale separation,

    G. B. De Luca and A. Tomasiello, “Leaps and bounds towards scale separation,” JHEP 12 (2021) 086, 2104.12773

    2021 arXiv

  45. [53]

    Cheeger bounds on spin-two fields,

    G. B. De Luca, N. De Ponti, A. Mondino, and A. Tomasiello, “Cheeger bounds on spin-two fields,”JHEP 12 (2021) 217, 2109.11560. 40

    2021 arXiv

  46. [54]

    Effective potential and warp factor dynamics,

    M. R. Douglas, “Effective potential and warp factor dynamics,”JHEP 03 (2010) 071, 0911.3378

    2010 arXiv

  47. [55]

    Aspects of Supergravity Theories,

    G. W. Gibbons, “Aspects of Supergravity Theories,” in XV GIFT Seminar on Supersymmetry and Supergravity. 6, 1984

    1984

  48. [56]

    Supergravity description of field theories on curved manifolds and a no-go theorem,

    J. M. Maldacena and C. Núñez, “Supergravity description of field theories on curved manifolds and a no-go theorem,”Int. J. Mod. Phys.A16 (2001) 822–855, hep-th/0007018

    2001 arXiv

  49. [57]

    An expert’s guide to training physics-informed neural networks,

    S. Wang, S. Sankaran, H. Wang, and P. Perdikaris, “An expert’s guide to training physics-informed neural networks,”arXiv preprint arXiv:2308.08468(2023)

    2023 arXiv

  50. [58]

    Learning representations by back-propagating errors,

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,”nature 323 (1986), no. 6088, 533–536

    1986

  51. [59]

    On the momentum term in gradient descent learning algorithms,

    N. Qian, “On the momentum term in gradient descent learning algorithms,”Neural networks : the official journal of the International Neural Network Society12 1 (1999) 145–151

    1999

  52. [60]

    A comprehensive study of non- adaptive and residual-based adaptive sampling for physics-informed neural net- works,

    C. Wu, M. Zhu, Q. Tan, Y. Kartha, and L. Lu, “A comprehensive study of non- adaptive and residual-based adaptive sampling for physics-informed neural net- works,” Computer Methods in Applied Mechanics and Engineering 403 (2023) 115671

    2023

  53. [61]

    Statistical mechanics of deep learning,

    Y. Bahri, J. Kadmon, J. Pennington, S. S. Schoenholz, J. Sohl-Dickstein, and S. Ganguli, “Statistical mechanics of deep learning,”Annual Review of Condensed Matter Physics11 (2020), no. 1, 501–528

    2020

  54. [62]

    Hyperbolic four-manifolds,

    B. Martelli, “Hyperbolic four-manifolds,”arXiv preprint arXiv:1512.03661(2015)

    2015 arXiv

  55. [63]

    Vinberg,Geometry II: Spaces of Constant Curvature

    ˙E. Vinberg,Geometry II: Spaces of Constant Curvature. Encyclopaedia of mathe- matical sciences. Springer-Verlag, 1993

    1993

  56. [64]

    Foundations of hyperbolic manifolds,

    J. G. Ratcliffe, “Foundations of hyperbolic manifolds,”Graduate Texts in Mathe- matics (2019)

    2019

  57. [65]

    Topics on totally discontinuous groups,

    H. C. Wang, “Topics on totally discontinuous groups,”Symmetric spaces (Short Courses, Washington Univ., St. Louis, Mo., 1969–1970)8 (1972) 459–487

    1972

  58. [66]

    Counting hyperbolic mani- folds,

    M. Burger, T. Gelander, A. Lubotzky, and S. Mozes, “Counting hyperbolic mani- folds,” Geometric and Functional Analysis12 (2002), no. 6, 1161–1173. 41

    2002

  59. [67]

    W. P. Thurston,The geometry and topology of three-manifolds: With a preface by Steven P. Kerckhoff, vol. 27. American Mathematical Society, 2022

    2022

  60. [68]

    Besse,Einstein Manifolds

    A. Besse,Einstein Manifolds. Classics in Mathematics. Springer Berlin Heidelberg, 2007

    2007

  61. [69]

    On right-angled reflection groups in hyperbolic spaces,

    L. Potyagailo and E. Vinberg, “On right-angled reflection groups in hyperbolic spaces,” Commentarii mathematici helvetici80 (2005), no. 1, 63–73

    2005

  62. [70]

    Right-angled coxeter polytopes, hyperbolic six-manifolds, and a problem of siegel,

    B. Everitt, J. G. Ratcliffe, and S. T. Tschantz, “Right-angled coxeter polytopes, hyperbolic six-manifolds, and a problem of siegel,”Mathematische Annalen 354 (2012) 871–905

    2012

  63. [71]

    Poisson,A relativist’s toolkit: the mathematics of black-hole mechanics

    E. Poisson,A relativist’s toolkit: the mathematics of black-hole mechanics. Cam- bridge university press, 2004

    2004

  64. [72]

    S. M. Carroll,Spacetime and geometry. Cambridge University Press, 2019

    2019

  65. [73]

    G. B. De Luca and Y. T. A. Law, “,”Work in progress

  66. [74]

    Understanding and mitigat- ing gradient flow pathologies in physics-informed neural networks,

    S. Wang, Y. Teng, and P. Perdikaris, “Understanding and mitigat- ing gradient flow pathologies in physics-informed neural networks,” SIAM Journal on Scientific Computing 43 (2021), no. 5, A3055–A3081, https://doi.org/10.1137/20M1318043

    2021 doi

  67. [75]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Sing- hal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimensional domains,”Advances in neural infor- mation processing systems33 (2020) 7...

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.