Pith. sign in

REVIEW 3 major objections 5 minor 68 references

Neural networks recover the full Schwarzschild black hole from Einstein equations alone, then find trapped Petrov type-I vacuum metrics.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 06:55 UTC pith:X2EZAQ6W

load-bearing objection Solid PINN upgrade that recovers Penrose–Schwarzschild cleanly and produces interesting local type-I trapped candidates; the “novel black hole” claim is still local. the 3 major comments →

arxiv 2607.05489 v1 pith:X2EZAQ6W submitted 2026-07-06 gr-qc cs.LGhep-th

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

Tancredi Schettini Gherardini , Edward Hirst , Alexander George Stapleton This is my paper
classification gr-qc cs.LGhep-th PACS 04.20.Jb04.25.D-04.70.Bw02.60.Lj
keywords Physics-informed neural networksEinstein metricsSchwarzschild geometryPetrov type Itrapped surfacesPenrose coordinatesLorentzian PINNsblack-hole search
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper shows that an unsupervised neural network can solve the vacuum Einstein equations in Lorentzian signature and recover the classic Schwarzschild black hole without ever being shown its metric components. By sampling a compact Penrose diagram and embedding the sphere so that topology is automatic, the network learns a metric whose curvature, scale, and spherical symmetry match the known solution. The same architecture is then freed from symmetry and given a different target: algebraically general Petrov type-I solutions that still have regular horizons and trapped interiors. The result is a practical search engine for vacuum black-hole geometries that do not fit standard analytic ansätze. A sympathetic reader cares because explicit Einstein metrics remain scarce; a method that can propose new candidates from invariant and causal conditions opens a route to less symmetric black holes.

Core claim

After Lorentzian and embedding upgrades, the same unsupervised AInstein architecture recovers the maximally extended Schwarzschild geometry from vacuum Einstein, Weyl-scale, and SO(3) losses, and, when those losses are replaced by a Petrov-profile, horizon-anchor, and trapped-surface objective, produces algebraically general (Petrov type I) numerically Ricci-flat Lorentzian metrics that possess regular Schwarzschild-scale horizons and genuinely trapped interiors.

What carries the argument

An ambient neural metric on R² × R³, pulled back through the stereographic embedding of S² and Penrose coordinates, trained by differentiable losses for Einstein residual, Weyl/Petrov invariants, non-degeneracy, and the trapping scalar Ξ.

Load-bearing premise

Low residuals on the chosen curvature, speciality, and trapping losses, plus later checks on null expansions, are taken as enough to certify that the learned metrics are genuine smooth vacuum black-hole spacetimes rather than local numerical artefacts.

What would settle it

On an independent sample set covering exterior and interior, either the Einstein residual fails to stay near zero, the speciality index collapses back to 1, the horizon curvature is not finite and Schwarzschild-scale, or the interior null expansions are not both negative after exact fibre-normal recomputation.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper extends the AInstein PINN architecture from Riemannian Einstein metrics to Lorentzian signature. Topology is encoded by learning an ambient metric on R^{2}×R^{3} with Penrose coordinates on R^{2} and the S^{2} metric obtained by pullback of the standard embedding, so chart consistency is architectural rather than enforced by an overlap loss. The method is validated by unsupervised recovery of the maximally extended Schwarzschild geometry from vacuum Einstein, Weyl-scale, SO(3), and determinant losses motivated by Birkhoff–Jebsen, with independent posterior checks on the speciality index, Kretschmann scale, and cubic curvature ratio. The objective is then generalised to a Petrov speciality-index profile, a horizon Weyl anchor, and a trapped-surface (Ξ) constraint, producing numerically Ricci-flat, algebraically general (type I) Lorentzian metrics with regular Schwarzschild-scale horizons and genuinely trapped interiors on the sampled domain.

Significance. If the Schwarzschild recovery and the type-I black-hole search hold under the stated diagnostics, the work supplies a flexible unsupervised search engine for vacuum Lorentzian geometries that is complementary to elliptic stationary constructions and Cauchy evolution. Strengths that should be credited include: (i) a reproducible public codebase; (ii) progressive calibration from 2D Lorentzian Einstein metrics through supervised and unsupervised Schwarzschild benchmarks; (iii) independent posterior invariants (S, Kr^{6}/m^{2}, ρ) not used as training targets that match analytic Schwarzschild values to high precision (Table II, Figs. 4–5, 7); and (iv) explicit trapped-surface certification via Ξ and null expansions for the type-I black-hole runs (Table IV, Fig. 6). The embedding-based global S^{2} treatment is a clean architectural improvement over chart-overlap losses. The main scientific value is methodological: invariant-profile and causal constraints can be dialled into a PINN search for black-hole-like metrics where analytic ansätze are restrictive.

major comments (3)
  1. §IV.D, Table IV, and the certification paragraphs: the claim of “potentially novel general-type Lorentzian Einstein metrics with a genuinely trapped interior” (abstract; §V) is supported only by residuals and expansions on samples with ρ≤0.85 inside the compactified Penrose domain (§III.C). Nothing reported enforces or checks that the Einstein residual remains small near the conformal boundary or the singularity, that the metric approaches a regular asymptotic form at I±/i^{0}, or that the ambient metric extends smoothly off the training measure. The objects are therefore rigorously local numerical Einstein metrics with trapped S^{2} fibres on the sampled support. The manuscript should either (a) add diagnostics near the domain edge and a clear asymptotic/regularity statement, or (b) systematically weaken the language from “vacuum black-hole spacetimes” to “local numerically Ricci-flat m
  2. §III.D.4–5, Eqs. (47)–(52): the speciality-index target S_targ is an arbitrary quadrupolar profile centred near 2, not a known Ricci-flat solution. Combined with free loss-weight schedules, trapping margins (β0, β1), and the Gaussian horizon anchor, this makes “novel type-I black hole” a statement about the training objective rather than an identification of a recognised geometry. The paper should state explicitly that S_targ is a search prior, report sensitivity of the recovered metrics to the profile choice (or at least to a second profile), and avoid implying that low L_profile_P alone classifies a unique physical solution class.
  3. §IV.D and footnote 11: half the seeds realise future-trapped (black-hole) interiors and half the time-reversed anti-trapped branch, because L_trap constrains Ξ but not the orientation of (θℓ, θn). This is acknowledged but under-emphasised relative to the abstract’s “black hole” language. The main text and abstract should state the orientation ambiguity up front and report black-hole vs white-hole fractions as a primary result, not only in a footnote.
minor comments (5)
  1. Table II: unsupervised Einstein/Weyl/Killing losses are orders of magnitude below the supervised baseline; a short discussion of why supervised component MSE underperforms the geometric residual objective (expressivity near the excluded boundary, loss landscape) would help readers interpret the benchmark.
  2. Eq. (28) and §II.D: Ξ is exact for round fibres and “corrected only by departure from roundness, which we find to be small.” Quantify that departure (e.g. multipole moments of the induced metric on the fibres) in the main text or appendix rather than only the ~5% expansion shift.
  3. Figs. 7–9: component panels are dense; consider adding a single scalar residual map (e.g. |R_μν| or Einstein residual) over the Penrose diagram for each experiment to make Ricci-flatness easier to assess at a glance.
  4. Appendix C / Tables V–VII: hyperparameter tables are useful; cross-reference the exact GitHub commit or config file used for the reported seeds so that the 10-run averages are fully reproducible.
  5. Typographical/notation: “AInstein Embedding Neural Networks” header (§III) vs “AINSTEIN”; consistent use of m vs mass parameter; arXiv date line “July 8, 2026” is fine for the preprint but should be checked against journal production.

Circularity Check

2 steps flagged

Mostly non-circular PINN validation/search; mild tautology only where training targets (S profile, horizon anchor, Ξ) are re-listed as independent invariants.

specific steps
  1. fitted input called prediction [§III.D.4–5; Tables III–IV (Re(S), √|I|r³, Khor, Ξint / trapped fraction)]
    "Lprofile_P := (1/N) Σ |27J(wa)²/(I(wa)³+ε_I) − S_targ(wa)|², S_targ(wa)≢1. ... Lhor := [Σ wa (r(wa)³√|Iθ|+ε − √3 m)²] / Σ ϕa ... Ltrap := (1/N) Σ [max(0, β(ra)−sign(ra−2m) Ξθ(wa))]². Table III/IV: Re(S) expected 2 → 1.994/1.993; √|I|r³/m → 1.750; Khor → 0.76; trapped fraction → 0.97–0.99."

    For type-I and BH runs the speciality index, horizon Weyl scale, and interior Ξ sign are explicit training targets. Re-reporting Re(S)≃2, √|I|r³≃√3 m, Khor≃0.75, and high trapped fraction as separate “invariant” success metrics is largely forced by those losses. Non-circular content remains the joint residual (especially LE,I) and non-target checks (det g, Killing residual, relative Frobenius distance to Schwarzschild).

  2. other [§II.B.2, §III.D.3, Eq. (42); Table II]
    "The architecture is first trained with the objective of recovering the Schwarzschild metric via losses encoding the vacuum Einstein equation, a quadratic Weyl scalar constraint, and the SO(3) symmetry of the resultant metric; directly motivated by the Birkhoff–Jebsen theorem. ... LW := (1/N) Σ (r(wa)³ √|Iθ(wa)|+ε − √3 m)². Here r(wa) is the analytic areal radius of the sampled Penrose point."

    Schwarzschild recovery is largely by design: losses encode the classical characterisation (vacuum + spherical symmetry + curvature scale), and LW uses the analytic Schwarzschild areal radius on the Penrose chart. This is transparent validation, not a hidden self-definition of a new law; posterior S, K, and ρ (not trained) still provide independent corroboration. Mild “expected success” only, not a closed circular derivation.

full rationale

This is a numerical methods paper: it encodes geometric constraints as PINN losses and reports that the optimiser finds metrics with small residuals. That is not a circular derivation of a physical law. For Schwarzschild, Ricci-flatness + SO(3) + a Weyl scale are deliberately the Birkhoff–Jebsen hypotheses; success is therefore expected if the network works, but the paper is explicit about this motivation and still checks posterior invariants (S, Kr^6/m^2, ρ) that were not training targets (speciality multiplier 0 in the Schwarzschild config). Birkhoff–Jebsen and the classical Schwarzschild invariants are external, not author uniqueness theorems. The type-I and black-hole runs are searches under chosen losses (ad-hoc S_targ≢1, Lhor, Ltrap), not predictions of a known closed-form solution; simultaneous low Einstein residual with non-Schwarzschild structure and trapping is non-tautological multi-objective content. Mild circularity appears only in tables that list Re(S)≈2, horizon √|I|r^3, Khor, and trapped fraction under “invariants” when those quantities were directly driven by Lprofile_P, Lhor, and Ltrap. Self-citation of AInstein is architectural background, not a load-bearing uniqueness chain. Score 2: one minor reporting tautology, central numerical claim still independent.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claims rest on standard vacuum GR, classical uniqueness theorems used only as targeting devices, and a collection of numerical free parameters (loss weights, speciality profile, trapping margins, network size, sampling density). No new physical fields or dimensions are postulated; the only ‘invented’ objects are architectural choices whose independent evidence is the numerical success itself.

free parameters (5)
  • speciality-index target profile S_targ
    Arbitrary quadrupolar profile centred near Re(S)=2, chosen by hand to repel from the type-D value; not derived from a known exact solution.
  • loss-weight schedule α_E, α_P, α_hor, α_trap, α_det
    Time-dependent multipliers that balance Einstein, Petrov, horizon, trapping and determinant terms; values listed in Tables V–VII and warmed up by hand.
  • trapping margins (β0, β1) and weight α_trap
    Two reported pairs (0.05,15) and (0.08,25) with β1=1; control the depth and strength of the interior trapping penalty.
  • horizon Gaussian width σ and mass parameter m
    m fixed to 1; σ sets the radial localisation of the curvature anchor about r=2m.
  • network architecture (width 256, 6 layers, GELU) and sampling (ρ=0.85, density powers)
    Hyper-parameters that define the function class and the training measure; chosen for numerical stability rather than derived.
axioms (5)
  • domain assumption Vacuum Einstein equation R_μν = 0 (λ=0) on a smooth Lorentzian 4-manifold
    Core PDE residual used in every unsupervised loss (§II.A, §III.D).
  • domain assumption Birkhoff–Jebsen theorem: spherically symmetric vacuum solutions are locally Schwarzschild
    Motivates the SO(3) + Einstein + Weyl-scale objective that targets the Schwarzschild branch (§II.B.2).
  • standard math Petrov speciality index S = 27 J² / I³ classifies algebraic type; S=1 for type D
    Standard invariant used both as posterior diagnostic and as training repeller (§II.B.1).
  • domain assumption A closed surface is (future) trapped when both null expansions θ_ℓ, θ_n < 0; Ξ = g(∇R,∇R) is a smooth proxy
    Quasi-local black-hole certificate of Penrose et al., implemented via automatic differentiation of areal radius (§II.D).
  • standard math Stereographic embedding of S² ⊂ ℝ³ induces the round metric by pullback and automatically glues the two charts
    Topological device that replaces the overlap loss of the original AInstein paper (§II.C, §III.B).

pith-pipeline@v1.1.0-grok45 · 29967 in / 3347 out tokens · 39670 ms · 2026-07-11T06:55:01.984258+00:00 · methodology

0 comments
read the original abstract

The AInstein architecture introduced an unsupervised neural method for solving the Riemannian Einstein equations on arbitrary manifolds. This Physics Informed Neural Network approach (PINN) is extended here to Lorentzian signature, validated by recovering the maximally extended Schwarzschild geometry, and tested as novel search method for arbitrary black hole solutions. The topology is built into the architecture by treating $S^{2}$ globally through its standard embedding, such that the network learns an ambient metric on the manifold $\mathbb{R}^{2} \times \mathbb{R}^{3}$, where Penrose coordinates are chosen for $\mathbb{R}^2$ and the metric on $S^{2}$ is obtained by pullback. The architecture is first trained with the objective of recovering the Schwarzschild metric via losses encoding the vacuum Einstein equation, a quadratic Weyl scalar constraint, and the $SO(3)$ symmetry of the resultant metric; directly motivated by the Birkhoff--Jebsen theorem. Following this, the objective is generalised to use the Petrov speciality index, a horizon curvature anchor, and a trapped-surface constraint, to allow search for algebraically general Petrov type I solutions, finding potentially novel general-type Lorentzian Einstein metrics with a genuinely trapped interior.

Figures

Figures reproduced from arXiv: 2607.05489 by Alexander George Stapleton, Edward Hirst, Tancredi Schettini Gherardini.

Figure 1
Figure 1. Figure 1: FIG. 1: Compactified Penrose domain used in Equation (4). The horizontal and vertical axes are the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: The sampling distribution, used in training and testing, shown for 2000 points, across the (a) [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Visualisations of the [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Diagonal metric components of the Penrose spacetime submanifold, for the analytic Schwarzschild [PITH_FULL_IMAGE:figures/full_fig_p026_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Invariant constant residuals for a representative Schwarzschild run. The three panels compare the [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Location of the trapped region for an example learnt black-hole model. [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7: Full component validation for the Schwarzschild run. The left block shows the learned metric [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Full component validation for the Petrov type-I search. The left block shows the learned metric [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Full component validation for the Petrov type-I black-hole search, with [PITH_FULL_IMAGE:figures/full_fig_p028_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

68 extracted references · 33 linked inside Pith

  1. [1]

    Petrov Classes, Weyl Invariants, and Curvature Scales 6

  2. [2]

    Spacetime Embedding 8 D

    Killing Vectors and Spherical Symmetry 7 C. Spacetime Embedding 8 D. Trapped Surfaces 9 III. AInstein Embedding Neural Networks 10 A. Warm up – Local Lorentzian Schwarzschild 10 B. Network Architecture 10 C. Data Generation 11 D. Training Objectives 12

  3. [3]

    Supervised Network 12

  4. [4]

    Unsupervised Local Lorentzian Network 12

  5. [5]

    Unsupervised Schwarzschild Network 12

  6. [6]

    Unsupervised Petrov Type-I Network 14

  7. [7]

    Results 16 A

    Unsupervised Petrov Type-I Black Hole Network 15 IV. Results 16 A. Local Lorentzian Searches 16 B. Schwarzschild Search 16 C. Petrov-Type I Vacuum Search 19 D. Petrov-Type I Black Hole Search 20 V. Conclusions and Outlook 21 Data Availability 22 Acknowledgements 22 A. Invariant Derivations for the Schwarzschild Solution 23 B. Component-Level Metric and Ri...

  8. [8]

    A classifier of the type may be defined through the complex speciality index ([40, 41])

    Petrov Classes, Weyl Invariants, and Curvature Scales Any smooth region of a 4D classical black-hole spacetime has a local Petrov type. A classifier of the type may be defined through the complex speciality index ([40, 41]). LetCabcd denote the Weyl tensor and define the self-dual Weyl tensor Cabcd :=C abcd −i ⋆Cabcd where ⋆Cabcd := 1 2 ϵabef Cef cd.(10) ...

  9. [9]

    The relevant infinitesimal sym- metries are the three rotations of the round two-sphere

    Killing Vectors and Spherical Symmetry The analytic Schwarzschild solution is spherically symmetric. The relevant infinitesimal sym- metries are the three rotations of the round two-sphere. Equivalently, the Killing algebra of the roundS 2 isso(3), generated by the infinitesimal rotations inherited from the ambient Euclidean space ([45]). In the Cartesian...

    2000

  10. [10]

    The base network still outputs the ambient Cholesky vector, while a wrapper applies Equation (32) and returns the flattened intrinsic metric

    Supervised Network The supervised model, where training instead approximates a known construction directly, is used for initialisation and benchmarking. The base network still outputs the ambient Cholesky vector, while a wrapper applies Equation (32) and returns the flattened intrinsic metric. For training sampleswa and target metricgtar µν, the default s...

  11. [11]

    The network described in Section IIIA is then trained with thisLE,Local loss only, forλ∈ {+1,0,−1}

    Unsupervised Local Lorentzian Network For the local two-dimensional calibration runs, for each training samplewa, define an Einstein error tensor Eµν(wa) :=R µν[g(θ)](wa)−λg (θ) µν (wa),(38) which is used to define the Einstein loss LE,Local := 1 N NX a=1 Eµν g(θ) E µα g(θ) E νβ Eαβ wa (39) whereg E is theEuclideanisedvariant of the metric 5. The network ...

  12. [12]

    The default Schwarzschild Lorentzian Einstein loss is then defined as 5 By this we meangE is obtained by the same vielbein asg, but usingδas the flat metric, instead ofη

    Unsupervised Schwarzschild Network For the four-dimensional Schwarzschild and general Petrov type searches, the vacuum equation isR µν = 0, soλ= 0is set 6. The default Schwarzschild Lorentzian Einstein loss is then defined as 5 By this we meangE is obtained by the same vielbein asg, but usingδas the flat metric, instead ofη. 6 Note the code also has funct...

  13. [13]

    A complementary search can instead use the speciality index in the loss function as a repeller from the Schwarzschild/type-DvalueS= 1

    Unsupervised Petrov Type-I Network The Schwarzschild experiment above uses the quadratic Weyl loss to fix the curvature scale and then checks the Petrov speciality index and Schwarzschild cubic invariant a posteriori. A complementary search can instead use the speciality index in the loss function as a repeller from the Schwarzschild/type-DvalueS= 1. This...

  14. [14]

    MathematicsandMachineLearningProgram

    Unsupervised Petrov Type-I Black Hole Network The type-I search of Section IIID4 drives the geometry away from the algebraically special locus while keeping it numerically Ricci-flat, but it fixes neither an absolute curvature scale nor any causal structure. The homothety-invariant lossLE,I is insensitive to constant rescalingsg7→Λ 2g, so the resulting ty...

    2000

  15. [15]

    Witten (Wiley, New York, 1962) pp

    R.Arnowitt, S.Deser, andC.W.Misner,in Gravitation: An Introduction to Current Research, edited by L. Witten (Wiley, New York, 1962) pp. 227–265, arXiv:gr-qc/0405109

    Pith/arXiv arXiv 1962

  16. [16]

    G. B. Cook, Living Reviews in Relativity3, 5 (2000), arXiv:gr-qc/0007085

    Pith/arXiv arXiv 2000

  17. [17]

    Lehner, Classical and Quantum Gravity18, R25 (2001), arXiv:gr-qc/0106072

    L. Lehner, Classical and Quantum Gravity18, R25 (2001), arXiv:gr-qc/0106072

    Pith/arXiv arXiv 2001

  18. [18]

    Shibata and T

    M. Shibata and T. Nakamura, Physical Review D52, 5428 (1995)

    1995

  19. [19]

    T. W. Baumgarte and S. L. Shapiro, Physical Review D59, 024007 (1999), arXiv:gr- qc/9810065

    arXiv 1999

  20. [20]

    Pretorius, Phys

    F. Pretorius, Phys. Rev. Lett.95, 121101 (2005), arXiv:gr-qc/0507014

    Pith/arXiv arXiv 2005

  21. [21]

    Campanelli, C

    M. Campanelli, C. O. Lousto, P. Marronetti, and Y. Zlochower, Physical Review Letters96, 111101 (2006), arXiv:gr-qc/0511048

    Pith/arXiv arXiv 2006

  22. [22]

    J. G. Baker, J. Centrella, D.-I. Choi, M. Koppitz, and J. van Meter, Physical Review Letters 96, 111102 (2006), arXiv:gr-qc/0511103

    Pith/arXiv arXiv 2006

  23. [23]

    J. M. Centrella, J. G. Baker, B. J. Kelly, and J. R. van Meter, Reviews of Modern Physics 82, 3069 (2010), arXiv:1010.5260 [gr-qc]

    Pith/arXiv arXiv 2010

  24. [24]

    M. A. Scheel, M. Boyle, T. Chu, L. E. Kidder, K. D. Matthews, and H. P. Pfeiffer, Physical Review D79, 024003 (2009), arXiv:0810.1767 [gr-qc]

    Pith/arXiv arXiv 2009

  25. [25]

    Löffler, J

    F. Löffler, J. Faber, E. Bentivegna, T. Bode, P. Diener, R. Haas, I. Hinder, B. C. Mundim, C. D. Ott, E. Schnetter, G. Allen, M. Campanelli, and P. Laguna, Classical and Quantum Gravity29, 115001 (2012), arXiv:1111.3344 [gr-qc]

    Pith/arXiv arXiv 2012

  26. [26]

    Wiseman, Class

    T. Wiseman, Class. Quant. Grav.20, 1137 (2003), arXiv:hep-th/0209051

    Pith/arXiv arXiv 2003

  27. [27]

    Headrick, S

    M. Headrick, S. Kitchen, and T. Wiseman, Class. Quant. Grav.27, 035002 (2010), arXiv:0905.1822 [gr-qc]

    Pith/arXiv arXiv 2010

  28. [28]

    A. Adam, S. Kitchen, and T. Wiseman, Class. Quant. Grav.29, 165002 (2012), arXiv:1105.6347 [gr-qc]

    Pith/arXiv arXiv 2012

  29. [29]

    O. J. C. Dias, J. E. Santos, and B. Way, Classical and Quantum Gravity33, 133001 (2016), arXiv:1510.02804 [hep-th]

    Pith/arXiv arXiv 2016

  30. [30]

    O. J. C. Dias, G. T. Horowitz, and J. E. Santos, Class. Quant. Grav.32, 145003 (2015), arXiv:1501.06574 [gr-qc]

    Pith/arXiv arXiv 2015

  31. [31]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, J. Comput. Phys.378, 686 (2019), arXiv:1711.10561 [cs.AI]

    Pith/arXiv arXiv 2019

  32. [32]

    Hirst, T

    E. Hirst, T. S. Gherardini, and A. G. Stapleton, AI for Science1, 025001 (2025)

    2025

  33. [33]

    A Machine Learning Approach to the Nirenberg Problem,

    G. Cortés, M. Esteban-Casadevall, Y. Feng, J. Henkel, E. Hirst, T. S. Gherardini, and A. G. Stapleton, “A Machine Learning Approach to the Nirenberg Problem,” (2026), arXiv preprint, arXiv:2602.12368 [cs.LG]

    arXiv 2026

  34. [34]

    Ashmore and Y.-H

    A. Ashmore and Y.-H. He, Fortschr. Phys.68, 2000068 (2020), arXiv:1910.08605 [hep-th]

    Pith/arXiv arXiv 2020

  35. [35]

    Douglas, S

    M. Douglas, S. Lakshminarasimhan, and Y. Qi, in Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference, Pro- ceedings of Machine Learning Research, Vol. 145, edited by J. Bruna, J. Hesthaven, and L. Zdeborova (PMLR, 2022) pp. 223–252

    2022

  36. [36]

    Larfors, A

    M. Larfors, A. Lukas, F. Ruehle, and R. Schneider, Mach. Learn. Sci. Tech.3, 035014 (2022), arXiv:2111.01436 [hep-th]

    Pith/arXiv arXiv 2022

  37. [37]

    Gerdes and S

    M. Gerdes and S. Krippendorf, Mach. Learn. Sci. Tech.4, 025031 (2023), arXiv:2211.12520 [hep-th]

    Pith/arXiv arXiv 2023

  38. [38]

    Berglund, G

    P. Berglund, G. Butbaia, T. Hüubsch, V. Jejjala, D. Mayorga Peña, C. Mishra, and J. Tan, Adv. Theor. Math. Phys.27, 1107 (2023), arXiv:2211.09801 [hep-th]. 31

    Pith/arXiv arXiv 2023

  39. [39]

    Harmonic1-forms on real loci of Calabi-Yau manifolds,

    M. R. Douglas, D. Platt, Y. Qi, and R. Barbosa, “Harmonic1-forms on real loci of Calabi-Yau manifolds,” (2024), arXiv preprint, arXiv:2405.19402 [math.DG]

    Pith/arXiv arXiv 2024

  40. [40]

    Neural and numerical methods for G2-structures on contact Calabi-Yau 7-manifolds,

    E. Heyes, E. Hirst, H. N. S. Earp, and T. S. R. Silva, “Neural and numerical methods for G2-structures on contact Calabi-Yau 7-manifolds,” (2026), arXiv preprint, arXiv:2602.12438 [math.DG]

    arXiv 2026

  41. [41]

    Machine learning gravity compactifications on negatively curved manifolds,

    G. B. De Luca, “Machine learning gravity compactifications on negatively curved manifolds,” (2025), arXiv:2501.00093 [hep-th]

    Pith/arXiv arXiv 2025

  42. [42]

    Minimal surfaces, knots, and neural networks,

    T. S. Gherardini and M. Usula, “Minimal surfaces, knots, and neural networks,” (2026), arXiv:2605.26234 [math.DG]

    Pith/arXiv arXiv 2026

  43. [43]

    Kumar, T

    P. Kumar, T. Mandal, and S. Mondal, JHEP10, 107 (2023), arXiv:2306.14817 [hep-th]

    Pith/arXiv arXiv 2023

  44. [44]

    Calculating Quasi-Normal Modes of Schwarzschild Black Holes with Physics Informed Neural Networks,

    N. Patel, A. Aykutalp, and P. Laguna, “Calculating Quasi-Normal Modes of Schwarzschild Black Holes with Physics Informed Neural Networks,” (2024), arXiv preprint, arXiv:2401.01440 [gr-qc]

    Pith/arXiv arXiv 2024

  45. [45]

    Einstein Fields: A Neural Per- spective To Computational General Relativity,

    S. S. Cranganore, A. Bodnar, A. Berzins, and J. Brandstetter, “Einstein Fields: A Neural Per- spective To Computational General Relativity,” (2025), arXiv preprint. Published as NeurIPS conference paper., arXiv:2507.11589 [cs.LG]

    arXiv 2025

  46. [46]

    He and J

    Y.-H. He and J. M. Pérez Ipiña, Phys. Lett. B832, 137213 (2022), arXiv:2201.01644 [gr-qc]

    Pith/arXiv arXiv 2022

  47. [47]

    Jejjala, S

    V. Jejjala, S. Mondkar, A. Mukhopadhyay, and R. Raj, Phys. Rev. D111, 026016 (2025), arXiv:2312.08442 [hep-th]

    Pith/arXiv arXiv 2025

  48. [48]

    Machine learning automorphic forms for black holes,

    V. Jejjala, S. Nampuri, D. Nxumalo, P. Roy, and A. Swain, “Machine learning automorphic forms for black holes,” (2025), arXiv preprint, arXiv:2505.05549 [hep-th]

    Pith/arXiv arXiv 2025

  49. [49]

    Hashimoto, K

    K. Hashimoto, K. Kyo, M. Murata, G. Ogiwara, and N. Tanahashi, Mach. Learn. Sci. Tech. 7, 015013 (2026), arXiv:2509.10866 [hep-th]

    arXiv 2026

  50. [50]

    Minimising Willmore Energy via Neural Flow,

    E. Hirst, H. N. S. Earp, and T. S. R. Silva, “Minimising Willmore Energy via Neural Flow,” (2026), arXiv preprint, arXiv:2604.04321 [math.DG]

    Pith/arXiv arXiv 2026

  51. [51]

    PINNs in More General Geometry,

    E. Hirst, “PINNs in More General Geometry,” (2026), arXiv preprint, arXiv:2604.25020 [math.DG]

    Pith/arXiv arXiv 2026

  52. [52]

    Penrose, Physical Review Letters10, 66 (1963)

    R. Penrose, Physical Review Letters10, 66 (1963)

    1963

  53. [53]

    The construction and application of penrose diagrams, with a focus on the maxi- mally analytically extended schwarzschild spacetime,

    C. Röken, “The construction and application of penrose diagrams, with a focus on the maxi- mally analytically extended schwarzschild spacetime,” (2026), arXiv:2507.23514 [gr-qc]

    arXiv 2026

  54. [54]

    Baker and M

    J. Baker and M. Campanelli, Physical Review D62, 127501 (2000), gr-qc/0003031

    Pith/arXiv arXiv 2000

  55. [55]

    Stephani, D

    H. Stephani, D. Kramer, M. MacCallum, C. Hoenselaers, and E. Herlt, Exact Solutions of Einstein’s Field Equations, 2nd ed. (Cambridge University Press, 2003)

    2003

  56. [56]

    Cherubini, D

    C. Cherubini, D. Bini, S. Capozziello, and R. Ruffini, International Journal of Modern Physics D11, 827 (2002), gr-qc/0302095

    Pith/arXiv arXiv 2002

  57. [57]

    Rosato, H

    N. Rosato, H. Nakano, and C. O. Lousto, Physical Review D104, 044047 (2021)

    2021

  58. [58]

    Zakhary and C

    E. Zakhary and C. B. G. Mcintosh, General Relativity and Gravitation29, 539–581 (1997)

    1997

  59. [59]

    J. M. Lee,Introduction to Riemannian Manifolds, 2nd ed. (Springer, 2018)

    2018

  60. [60]

    G. D. Birkhoff, Relativity and Modern Physics (Harvard University Press, Cambridge, MA, 1923)

    1923

  61. [61]

    J. T. Jebsen, Arkiv för Matematik, Astronomi och Fysik15, 1 (1921)

    1921

  62. [62]

    Israel, Physical Review164, 1776 (1967)

    W. Israel, Physical Review164, 1776 (1967)

    1967

  63. [63]

    Penrose, Phys

    R. Penrose, Phys. Rev. Lett.14, 57 (1965)

    1965

  64. [64]

    S. A. Hayward, Phys. Rev. D49, 6467 (1994)

    1994

  65. [65]

    Ashtekar and B

    A. Ashtekar and B. Krishnan, Living Rev. Rel.7, 10 (2004), arXiv:gr-qc/0407042

    Pith/arXiv arXiv 2004

  66. [66]

    C. W. Misner and D. H. Sharp, Phys. Rev.136, B571 (1964). 32

    1964

  67. [67]

    Quevedo, S

    H. Quevedo, S. Toktarbay, and A. Yerlan, Version published in International Journal of Mathematics and Physics3, 133 (2012), arXiv:1310.5339 [gr-qc]

    Pith/arXiv arXiv 2012

  68. [68]

    Newman and R

    E. Newman and R. Penrose, J. Math. Phys.3, 566 (1962). 33

    1962