Pith. sign in

REVIEW 4 major objections 5 minor 46 references

This paper builds a complete numerical pipeline for warped Type IIB compactifications on Calabi-Yau threefolds and shows that near a conifold about 0.5 percent of the volume is a strongly warped throat.

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

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

T0 review · deepseek-v4-flash

2026-08-01 15:29 UTC pith:6APNJ2PV

load-bearing objection First complete numerical GKP warp pipeline on a compact CY3, well-checked and honest; the 0.5% throat number is a feature of the smeared-source toy, not a real Type IIB vacuum. the 4 major comments →

arxiv 2607.18402 v1 pith:6APNJ2PV submitted 2026-07-20 hep-th

Warped Numerical Calabi-Yau Metrics

Severin L\"ust , Fabian Ruehle , Simon Schreyer This is my paper
classification hep-th
keywords Calabi-Yau metricswarped flux compactificationsphysics-informed neural networksconifold throatsimaginary self-dual fluxsingular bulk problemnumerical Ricci-flat metricswarp factor
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.

The paper tries to establish that all three ingredients of a warped Type IIB flux compactification—the Ricci-flat Calabi-Yau metric, the imaginary self-dual three-form flux, and the warp factor solving a sourced Poisson equation—can be computed numerically on a compact Calabi-Yau threefold using physics-informed neural networks. It demonstrates this on a one-parameter quintic family at two flux vacua, one near the conifold locus and one far away as a cross-check. The concrete payoff is a first quantitative handle on the singular bulk problem: for the near-conifold benchmark, roughly 0.5 percent of the total Calabi-Yau volume sits in the strongly warped throat, where the warp factor is an order of magnitude larger than in the bulk. A skeptical but sympathetic reader cares because warping is what generates hierarchies and controls moduli stabilization, and until now the local distribution of warping on an explicit compact manifold was essentially unknown. The paper also reports several reusable technical improvements in point sampling, network architecture, and loss design.

Core claim

On its own terms, the paper's central claim is that the full data of a warped Type IIB flux background—the Ricci-flat Calabi-Yau metric, the harmonic (2,1)-forms that encode imaginary self-dual flux, and the warp factor e^{-4A}—can be approximated together on a compact Calabi-Yau threefold by physics-informed neural networks. Applied to a one-parameter quintic family, the pipeline stabilizes two flux vacua from periods, one close to the conifold and one away from it. The near-conifold vacuum develops a clear tail in the warp-factor distribution that is absent in the cross-check vacuum, and summing the Monte-Carlo weights of the points in that tail gives the headline quantity: the strongly wa

What carries the argument

The work is carried by three coupled neural-network approximations. A phi-network corrects the ambient Kähler potential so that the induced metric satisfies the Monge-Ampere equation, giving the Ricci-flat Calabi-Yau metric. A second network learns a correction to the reference (2,1)-form so that it is harmonic with respect to that metric, making the three-form flux imaginary self-dual. A third network solves the sourced Poisson equation for e^{-4A} by minimizing its residual, with a weighted robust loss that up-weights rare, strongly negative source points—essential for making the throat visible. An improved point-sampling algorithm that reweights toward the Calabi-Yau volume form provides

Load-bearing premise

The physical conclusion rests on replacing the localized negative D3 charge of an orientifold with a smeared constant source, because the quintic family used here has no orientifold involution; if that source is not a faithful stand-in, the 0.5 percent throat-volume statement is a property of the toy equation rather than of a real Type IIB vacuum.

What would settle it

Repeat the same computation on a quintic threefold that admits a genuine orientifold involution, placing the negative charge on the orientifold locus instead of smearing it. If the strongly warped volume fraction moves far from 0.5 percent, or if the tail in the warp-factor distribution disappears, the paper's benchmark conclusion does not survive.

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

If this is right

  • The warped volume fraction is now a computable quantity: for the near-conifold benchmark it is about 0.5 percent, with an order-of-magnitude warp enhancement.
  • The pipeline can be rerun at different distances from the conifold, so the size and profile of the throat as a function of modulus stabilization can be mapped.
  • The cross-check vacuum away from the conifold shows no throat tail, confirming that the observed bump is tied to the conifold flux vacuum and not an artifact of the loss function.
  • The technical ingredients—curvature-adapted sampling, spectral features, multi-step training, and the weighted robust loss—transfer to other geometric PDE problems on Calabi-Yau manifolds.
  • Derived quantities such as the Euler number and the Weil-Petersson metric agree with exact or topological values away from the conifold, supporting trust in the warp-factor result in the controlled region.

Where Pith is reading between the lines

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

  • Because the paper uses a smeared rather than localized negative source, the 0.5 percent figure should be read as a method demonstration, not a prediction; a genuine orientifold source could concentrate warping differently and shift the throat fraction by an order of magnitude.
  • The paper's total throat volume is spread over 125 symmetric conifolds, implying each individual throat is only about 0.004 percent of the volume; resolving a single throat seems to require adaptive sampling targeted at the conifold.
  • The modular structure of the pipeline suggests a natural scan: fixing the metric and recomputing only the harmonic representative and warp factor for many flux quanta could turn the singular bulk problem into a statistical statement over flux space.

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

4 major / 5 minor

Summary. The paper presents a numerical pipeline that aims to approximate the three ingredients of a GKP-type warped Type IIB compactification on a Calabi-Yau threefold: the Ricci-flat metric, harmonic (2,1)-forms encoding ISD three-form flux, and the warp factor solving the sourced Poisson equation. The pipeline is applied to the Dwork quintic at two flux vacua, one near the conifold and one near the Landau-Ginzburg point. The authors report sub-percent agreement with the topological Euler number and the Weil-Petersson metric away from the conifold, and use the near-conifold vacuum to estimate that the strongly warped throat occupies about 0.5% of the CY volume with an order-of-magnitude warp enhancement. The paper also introduces several numerical techniques: improved point sampling, a spectral feature-engineered network, multi-step PINN training, and a weighted Huber loss.

Significance. If the physical conclusions were fully supported, this would be the first numerical quantification of the singular-bulk problem on a compact CY threefold and a useful blueprint for numerical GKP data. The methods are nontrivial, the code is public, and the metric/form part of the pipeline has genuine external anchors: sub-percent Euler number and Weil-Petersson checks away from the conifold, and a clean null result (no throat) for the LG cross-check. These are real strengths and make the paper potentially valuable as a methods contribution. However, the headline physical claim concerns a toy equation, not an actual Type IIB orientifold vacuum, and the warp-factor solution itself is not independently benchmarked. The central claim is therefore only partially defensible as stated.

major comments (4)
  1. [§4.3, Eq. (4.23); Abstract; §6.5.2] The warp-factor equation is not the GKP equation for a physical vacuum. In §4.3 the negative term in (4.23) is replaced by a smeared, constant negative energy because the Dwork family has no orientifold involution, so the localized O3/O7 charges of (2.8) are absent. Since this negative source is what cancels the D3 tadpole and shapes the solution, the ≈0.5% throat volume and order-of-magnitude warp enhancement reported in the Abstract and §6.5.2 are properties of a toy Poisson problem, not predictions for a Type IIB flux vacuum. The limitation is acknowledged in §4.3 and §7, but the Abstract's claim to compute 'all three ingredients required by the GKP setup' and the physical interpretation in §6.5.2 overstate what is done. Either a genuine orientifold projection with localized sources must be implemented, or the physical claims must be explicitly reframed as a controlled toy-model study
  2. [§3.1] The flux construction is formulated in H^3_-(X,Z), the orientifold-odd middle cohomology, and the flux scan is performed in this space. However, §4.3 states that the Dwork family does not admit an orientifold involution, and no involution is defined anywhere in the paper. Without an involution, the decomposition H^3 = H^3_+ ⊕ H^3_- is not defined, so the status of the 'flux vacua' in Table 1 as GKP vacua is unclear. At minimum, the paper should state that the H^3_- notation is only formal and that the resulting backgrounds are not orientifold vacua; ideally, a quintic family with a known orientifold action should be used.
  3. [§6.4] The manuscript states that the warp-factor solution 'cannot really be benchmarked' other than by a drop in the PDE residual. A falling training loss is not evidence of correctness, and the KL divergence between Methods 1 and 2 compares two approximate evaluations of the same equation using the same metric and the same smeared-source approximation. The throat-volume claim in §6.5.2 therefore rests on an unvalidated numerical solution. Please add an independent benchmark: for example, compare against a standard finite-difference/spectral solver for the same RHS on a subset of points, or against an analytic local conifold/KS solution, and report a convergence study in NN size, sample size, and hyperparameters.
  4. [§6.5.2; §5.5] The 0.5% throat-volume estimate is obtained by summing weights of points in a 'bump' in the CCDF. This procedure depends on mean alignment, on the identification of the bump, and on the weighted-Huber hyperparameters δ=0.3, α=10, c=5 introduced in §5.5. The paper shows single representative runs, and with only O(100) throat points spread over 125 conifolds the estimate has no quantified uncertainty. Please provide a robustness analysis over the 20 point sets, over hyperparameter choices, and over the definition of the CCDF threshold, and report the spread of the resulting volume fraction.
minor comments (5)
  1. [Table 1; §3.2] The value of ψ for the LG-point vacuum is printed as 0.5+0.1i in Table 1 but as 0.5+0.01i in the text of §3.2. Please reconcile.
  2. [§3.2] For the near-conifold vacuum, |Z| ≈ 0.17, so the solution is not exponentially close to the conifold. The text acknowledges this, but the phrase 'proper conical region' and the subsequent KS-throat discussion should be tempered accordingly.
  3. [References] Several bibliography entries contain unprocessed LaTeX or formatting artifacts (e.g., refs [9] and [22]) and should be cleaned before publication.
  4. [Fig. 7] The caption refers to 'dataset 4 and 8', but datasets are not defined in the text. Please define them or remove the reference.
  5. [§5.5, Eq. (5.9)] Eq. (5.9) writes Huber(y,f(x)), while Eq. (5.10) defines the Huber loss as a function of the residual |y-f(x)|. Please make the notation consistent.

Circularity Check

0 steps flagged

No significant circularity: the numerical pipeline is anchored by external benchmarks; the toy orientifold source is an acknowledged scope limitation, not a fitted input.

full rationale

The derivation chain is not circular. The CY metric is learned from the Monge-Ampere residual (4.3) and benchmarked against the topological Euler number chi=-200 (Tables 2 and 3) and against the period-derived Weil-Petersson metric (Table 2). The harmonic (2,1)-forms are obtained by imposing the harmonicity conditions (4.13)-(4.15) on a closed reference representative (4.8), with no target data being fed into the network. The warp factor solves the Poisson equation (4.23); its source is computed from flux coefficients (3.11)-(3.12) and from Omega/chi normalizations matched to period data, and the network is trained only by minimizing the PDE residual (5.9)-(5.11). No fitted parameter is renamed as a prediction: the 0.5% throat volume is obtained by integrating the tail of the solved warp-factor CCDF, not by construction from the loss or from the source term. The main caveat is an acknowledged modeling limitation rather than circularity: the Introduction states 'our study was performed on a CY which does not have an orientifold action at the point in moduli space we studied', Section 4.3 says 'we mimic the orientifold by a smeared, constant negative energy', and the Conclusions reiterate that 'we therefore had to mimic the negative tension by a constant smeared source.' Moreover, Section 6.4 notes that the warp-factor solution 'cannot really be benchmarked ... other than by noting a drop in the error of the Poisson equation.' These passages make the physical throat-size result conditional on a toy source, an external-validity/scope concern, not a self-referential reduction. Self-citations ([6,7,29,30]) are background references on KKLT and warping and are not load-bearing for the numerical pipeline. Score 1 reflects the absence of circular steps while acknowledging the unbenchmarked, toy-source status of the headline physical numbers.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 1 invented entities

The pipeline rests on standard supergravity and complex-geometry axioms, but the physical conclusion adds one ad hoc entity (smeared negative energy) and several hand-set numerical parameters (Huber weights, flux scan range, normalization matching, mean alignment). These are the main reasons the 0.5% throat volume is not yet a robust string-theory prediction.

free parameters (5)
  • Flux scan range Nmax = 2
    Flux quanta are scanned in {-2,...,2}; restricts the set of vacua and the benchmark choices (§3.2).
  • Huber loss threshold δ = 0.3
    Chosen by hand in (5.10); controls how outliers are treated in the warp-factor loss.
  • Huber weighting α, c = α=10, c=5
    Weights in (5.11) are chosen to saturate rare negative sources and directly shape the CCDF bump used for the 0.5% throat-volume estimate.
  • CY volume / Ω normalization matching factors = matched separately to period computation
    §4.3: Ω and χ normalizations are rescaled to match period-derived values; the two factors differ slightly because the numerics are not exact.
  • Warp factor integration constant / mean alignment = not fixed
    §6.5: e^{-4A} is determined only up to a constant; distributions are mean-aligned, so order-of-magnitude statements are relative to a chosen normalization.
axioms (7)
  • domain assumption The ten-dimensional metric takes the GKP warped form (2.1) with constant axio-dilaton (2.9)
    Invoked in Section 2 as the starting point; violations (e.g., non-constant τ from D7/O7) are acknowledged but not modeled.
  • domain assumption Harmonic (2,1)-forms can be obtained by adding an exact correction Δm to the reference m from the ambient-space construction (Eq. 4.17)
    Section 4.2 assumes this representative exists and is found by the NN; no proof of convergence to the true harmonic representative is given.
  • standard math The Dwork quintic periods (3.15)-(3.18), expanded to order 50, determine the stabilized moduli via D_i W=0
    Uses known Candelas-de la Ossa-Green-Parkes periods; the truncation error is asserted to be below percent-level.
  • standard math Shiffman-Zelditch theorem: zeros of random sections are FS-uniform; λ-adjustment yields mass ≈1 points
    Section 5.1 builds the improved point sampling on this theorem.
  • ad hoc to paper A smeared constant negative energy source can mimic an orientifold's negative D3 charge in (2.7)
    Section 4.3: 'we mimic the orientifold by a smeared, constant negative energy'; this is a toy-model replacement, not a string-theoretic source.
  • domain assumption The Dwork family admits no orientifold involution
    Taken from [43] and used to justify the smeared source instead of actual O-planes.
  • domain assumption Warp factor is defined only up to an additive constant, so distributions can be mean-aligned for comparison (Frey-Torroba-Underwood-Douglas [46])
    Section 6.5 uses this invariance to compare NNs/methods; the same freedom means absolute warp ratios are not fixed.
invented entities (1)
  • Smeared constant negative D3-energy source (orientifold mimic) no independent evidence
    purpose: Serves as the negative ρ_loc3 term in the warp-factor Poisson equation (2.7) in a geometry without orientifold planes
    No O-plane or D7-brane is present in the Dwork family; its distribution and magnitude are chosen ad hoc, so the throat-volume result is not a property of a known string vacuum.

pith-pipeline@v1.3.0-alltime-deepseek · 22248 in / 13259 out tokens · 107930 ms · 2026-08-01T15:29:51.332277+00:00 · methodology

0 comments
read the original abstract

We compute numerical warped Type IIB flux backgrounds on Calabi-Yau threefolds following the construction of Giddings, Kachru, and Polchinski. Using physics-informed neural networks, we approximate all three ingredients required by the GKP setup: the Ricci-flat Calabi-Yau metric, the harmonic (2,1)-forms representing the imaginary self-dual three-form flux, and the warp factor, which solves a sourced Poisson equation on the internal manifold. We apply our pipeline to the Dwork family of quintics for two different flux vacua, one near a conifold point and one away from it, the latter serving as a numerical cross-check. With these tools, we study the singular bulk problem, and find that for our benchmark point close to the conifold, approximately 0.5 percent of the total Calabi-Yau volume sits in the throat, and the warp factor is an order of magnitude larger as compared to the bulk. We also introduce several improvements to techniques used for numerical studies of CY metrics and quantities derived from them that might be of interest independently of our application. These include an improved point sampling algorithm that produces samples that are more uniform under the Calabi-Yau measure, a feature-engineered spectral network for the metric, multi-step physics-informed training, and a weighted Huber loss tailored to stiff PDEs with highly non-uniform sources.

Figures

Figures reproduced from arXiv: 2607.18402 by Fabian Ruehle, Severin L\"ust, Simon Schreyer.

Figure 1
Figure 1. Figure 1: Distribution of points on a square torus lattice when sampled with 1,3, and 5 different metrics using point rejection. transition maps on the manifold. Moreover, since the NN is now a globally defined real function, it cannot change the K¨ahler class, and we can also drop the volume loss (although keeping or dropping it does not have a big impact on performance). This can be arranged easily: instead of usi… view at source ↗
Figure 2
Figure 2. Figure 2: Two-step training of the CY metric for two different model sizes (2 hidden layers with 64 neurons, and 4 hidden layers with 256 neurons). approximate the CY metric as the pullback of g˜ = ∂ ¯∂ [KFS + a1ϕ1 + a2ϕ2 + . . .] , (5.8) where the ai are real parameters controlling the strengths of the corrections. One then first trains ϕ1 (we use a1 = 1) until the loss does not improve significantly anymore. At th… view at source ↗
Figure 3
Figure 3. Figure 3: Training the CY metric ϕ NN for 30 epochs with ADAM, followed by 40 epochs with L-BFGS. a PINN is first trained with ADAM and then with a few steps of L-BFGS was already proposed in [21]. We implemented this strategy using scipy’s L-BFGS optimizer, where we trained a small NN (2 hidden layers, 32 neurons each) on 100k points close to the conifold for 30 epochs with ADAM, followed by 40 epochs of L-BFGS tra… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of NN architectures for approximating the harmonic representatives using IPS point sampling (left) and standard (right). For standard sampling, we fixed the activation to GeLU. for the benchmark point close to the conifold.6 We try models with 2 and 4 hidden layers with 64 and 256 neurons each. We also use layer normalization and look at RELU, GELU, SiLU activation functions. Finally, we compare… view at source ↗
Figure 5
Figure 5. Figure 5: 50 epochs of training the warp factor PINN (4 hidden layers, 256 neurons) with MSE loss, followed by 50 epochs of weighted Huber loss. want to quantify how “large” the region is in which the geometry becomes strongly warped. We discuss our results for different point samples and for training with and without the weighted Huber loss. As mentioned in the Introduction, we study a point in CY moduli space that… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of the CCDFs without (left) and with (right) weighted Huber loss training away from the conifold. 6.5.2 Close to conifold 0.00010 0.00015 0.00020 0.00025 0.00030 0.00035 0.00040 0.00045 e 4A phys (mean-aligned) 0 200 400 600 800 1000 count Warp factor distributions (means aligned), standard sampling method 1, NN: L4 N256 method 1, NN: L4 N256, Huber loss method 2, NN: L4 N256 method 2, NN: L4 N2… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of two warp-factor distributions using two-step metric training for our solutions away from and close to the conifold (dataset 4 and 8, respectively). The means of the distributions are aligned. Before repeating the CCDF analysis close to the conifold, we show the full warp factor distribution we obtain with and without Huber loss in [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Warp factor CCDF for dataset 6 with standard and IPS sampling for different NNs and computation methods, close to the conifold. 10 0 4 × 10 1 6 × 10 1 2 × 10 0 e 4A / e 4A 10 5 10 4 10 3 10 2 10 1 10 0 P(e 4A > x) CCDF of Warp Factor close to conifold, standard sampling method 1, NN: L2 N64 method 1, NN: L4 N256 method 2, NN: L2 N64 method 2, NN: L4 N256 10 0 10 1 e 4A / e 4A 10 5 10 4 10 3 10 2 10 1 10 0 … view at source ↗
Figure 9
Figure 9. Figure 9: Warp factor distribution before and after weighted Huber loss for standard sampling close to the conifold. should have a relatively large weight in the Monte-Carlo integration. Indeed, we find that for the points indicated to have strong warping by the CCDF, their integration weight is a factor of 5 larger than that of bulk points. Next, we compare again the KDE-base KL between the two computation methods.… 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

46 extracted references · 30 linked inside Pith

  1. [1]

    Hierarchies from fluxes in string com- pactifications,

    S. B. Giddings, S. Kachru, and J. Polchinski “Hierarchies from fluxes in string com- pactifications,”Phys. Rev. D66(2002) 106006 [hep-th/0105097]

    Pith/arXiv arXiv 2002

  2. [2]

    Supergravity and a confining gauge theory: Du- ality cascades and chi SB resolution of naked singularities,

    I. R. Klebanov and M. J. Strassler “Supergravity and a confining gauge theory: Du- ality cascades and chi SB resolution of naked singularities,”JHEP08(2000) 052 [hep-th/0007191]

    Pith/arXiv arXiv 2000

  3. [3]

    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]

    Pith/arXiv arXiv 2003

  4. [4]

    Candidate de Sitter vacua,

    L. McAllister, J. Moritz, R. Nally, and A. Schachner “Candidate de Sitter vacua,”Phys. Rev. D111(2025) no. 8, 086015 [2406.13751]

    Pith/arXiv arXiv 2025

  5. [5]

    Control issues of KKLT,

    X. Gao, A. Hebecker, and D. Junghans “Control issues of KKLT,”Fortsch. Phys.68 (2020) 2000089 [2009.03914]

    Pith/arXiv arXiv 2020

  6. [6]

    Holography and the KKLT scenario,

    S. L¨ ust, C. Vafa, M. Wiesner, and K. Xu “Holography and the KKLT scenario,”JHEP 10(2022) 188 [2204.07171]

    Pith/arXiv arXiv 2022

  7. [7]

    KKLT compactifications ex nihilo,

    I. Bena, Y. Li, and S. L¨ ust “KKLT compactifications ex nihilo,”Phys. Rev. D113 (2026) no. 4, 046013 [2410.22400]

    Pith/arXiv arXiv 2026

  8. [8]

    Moduli-dependent Calabi-Yau and SU(3)-structure metrics from Machine Learning,

    L. B. Anderson, M. Gerdes, J. Gray, S. Krippendorf, N. Raghuram, and F. Ruehle “Moduli-dependent Calabi-Yau and SU(3)-structure metrics from Machine Learning,” JHEP05(2021) 013 [2012.04656]

    Pith/arXiv arXiv 2021

  9. [9]

    Numerical Calabi-Yau metrics from holomorphic networks,

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

    Pith/arXiv arXiv 2012

  10. [10]

    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]

    Pith/arXiv arXiv 2020

  11. [11]

    Neural network approximations for Calabi-Yau metrics,

    V. Jejjala, D. K. Mayorga Pena, and C. Mishra “Neural network approximations for Calabi-Yau metrics,”JHEP08(2022) 105 [2012.15821]

    Pith/arXiv arXiv 2022

  12. [12]

    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,”Proceedings of Advances in Neural Information Processing Systems 34 - Machine Learning and the Physical Sciences](11, 2021) [2111.01436]. 31

    Pith/arXiv arXiv 2021

  13. [13]

    Numerical metrics for complete intersection and Kreuzer–Skarke Calabi–Yau manifolds,

    M. Larfors, A. Lukas, F. Ruehle, and R. Schneider “Numerical metrics for complete intersection and Kreuzer–Skarke Calabi–Yau manifolds,”Mach. Learn. Sci. Tech.3 (2022) no. 3, 035014 [2205.13408]

    Pith/arXiv arXiv 2022

  14. [14]

    cymetric,

    F. Ruehle and R. Schneider “cymetric,”GitHub(2022) [https://github.com/ruehlef/cymetric]

    2022

  15. [15]

    Machine-learned Calabi–Yau metrics and curvature,

    P. Berglund, G. Butbaia, T. H¨ ubsch, V. Jejjala, D. Mayorga Pe˜ na, C. Mishra, and J. Tan “Machine-learned Calabi–Yau metrics and curvature,”Adv. Theor. Math. Phys. 27(2023) no. 4, 1107–1158 [2211.09801]

    Pith/arXiv arXiv 2023

  16. [16]

    Physical Yukawa couplings in heterotic string compactifications,

    G. Butbaia, D. Mayorga Pe˜ na, J. Tan, P. Berglund, T. H¨ ubsch, V. Jejjala, and C. Mishra “Physical Yukawa couplings in heterotic string compactifications,”Adv. Theor. Math. Phys.28(2024) no. 8, 2783–2822 [2401.15078]

    Pith/arXiv arXiv 2024

  17. [17]

    Com- putation of quark masses from string theory,

    A. Constantin, C. S. Fraser-Taliente, T. R. Harvey, A. Lukas, and B. Ovrut “Com- putation of quark masses from string theory,”Nucl. Phys. B1010(2025) 116778 [2402.01615]

    Pith/arXiv arXiv 2025

  18. [18]

    cymyc: Calabi-Yau Metrics, Yukawas, and Curvature,

    G. Butbaia, D. Mayorga Pe˜ na, J. Tan, P. Berglund, T. H¨ ubsch, V. Jejjala, and C. Mishra “cymyc: Calabi-Yau Metrics, Yukawas, and Curvature,”JHEP03(2025) 028 [2410.19728]

    Pith/arXiv arXiv 2025

  19. [19]

    Quark masses and mixing in string-inspired models,

    A. Constantin, C. S. Fraser-Taliente, T. R. Harvey, L. T. Y. Leung, and A. Lukas “Quark masses and mixing in string-inspired models,”JHEP06(2025) 175 [2410.17704]

    Pith/arXiv arXiv 2025

  20. [20]

    Reproducing Standard Model fermion masses and mixing in string theory: A heterotic line bundle study,

    A. Constantin, L. T. Y. Leung, A. Lukas, and L. A. Nutricati “Reproducing Standard Model fermion masses and mixing in string theory: A heterotic line bundle study,” Phys. Rev. D113(2026) no. 4, 046005 [2507.03076]

    Pith/arXiv arXiv 2026

  21. [21]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,”Journal of Computational Physics378(2019) 686–707

    2019

  22. [22]

    Discovery of unstable singularities,

    Y. Wang, M. Bennani, J. Martens, S. Racani` ere, S. Blackwell, A. Matthews, S. Nikolov, G. Cao-Labora, D. S. Park, M. Arjovsky, D. Worrall, C. Qin, F. Alet, B. Kozlovskii, N. Tomaˇ sev, A. Davies, P. Kohli, T. Buckmaster, B. Georgiev, J. G´ omez-Serrano, R. Jiang, and C.-Y. Lai “Discovery of unstable singularities,” 2025

    2025

  23. [23]

    Non-uniqueness and symmetries for the nirenberg problem using computer assistance,

    D. Platt “Non-uniqueness and symmetries for the nirenberg problem using computer assistance,” 2026

    2026

  24. [24]

    A Machine Learning Approach to the Nirenberg Problem,

    G. Cort´ es, M. Esteban-Casadevall, Y. Feng, J. Henkel, E. Hirst, T. S. Gherardini, and A. G. Stapleton “A Machine Learning Approach to the Nirenberg Problem,” [2602.12368]

    arXiv

  25. [25]

    Harmonic 1-forms on real loci of Calabi-Yau manifolds,

    M. R. Douglas, D. Platt, Y. Qi, and R. Barbosa “Harmonic 1-forms on real loci of Calabi-Yau manifolds,” 5, 2024. 32

    2024

  26. [26]

    AInstein: Numerical Einstein Metrics via Machine Learning,

    E. Hirst, T. S. Gherardini, and A. G. Stapleton “AInstein: Numerical Einstein Metrics via Machine Learning,”AI Sci.1(2025) no. 2, 025001 [2502.13043]

    arXiv 2025

  27. [27]

    PINNs in More General Geometry,

    E. Hirst “PINNs in More General Geometry,” [2604.25020]

    Pith/arXiv arXiv

  28. [28]

    Uplifting Runaways,

    I. Bena, E. Dudas, M. Gra˜ na, and S. L¨ ust “Uplifting Runaways,”Fortsch. Phys.67 (2019) no. 1-2, 1800100 [1809.06861]

    Pith/arXiv arXiv 2019

  29. [29]

    Effective Theory of Warped Compactifications and the Impli- cations for KKLT,

    S. L¨ ust and L. Randall “Effective Theory of Warped Compactifications and the Impli- cations for KKLT,”Fortsch. Phys.70(2022) no. 7-8, 2200103 [2206.04708]

    Pith/arXiv arXiv 2022

  30. [30]

    Effective potentials, warping, and implications for F-term uplifting,

    A. Hebecker, S. L¨ ust, A. Schachner, and S. Schreyer “Effective potentials, warping, and implications for F-term uplifting,”JHEP05(2026) 286 [2512.17995]

    arXiv 2026

  31. [31]

    Gaugino condensation and small uplifts in KKLT,

    F. Carta, J. Moritz, and A. Westphal “Gaugino condensation and small uplifts in KKLT,”JHEP08(2019) 141 [1902.01412]

    Pith/arXiv arXiv 2019

  32. [32]

    Numerical Weil–Petersson metrics on moduli spaces of Cal- abi–Yau manifolds,

    J. Keller and S. Lukic “Numerical Weil–Petersson metrics on moduli spaces of Cal- abi–Yau manifolds,”J. Geom. Phys.92(2015) 252–270 [0907.1387]

    Pith/arXiv arXiv 2015

  33. [33]

    Distribution of zeros of random and quantum chaotic sections of positive line bundles,

    B. Shiffman and S. Zelditch “Distribution of zeros of random and quantum chaotic sections of positive line bundles,”Communications in Mathematical Physics200(1999) 661–683

    1999

  34. [34]

    Multi-stage neural networks: Function approximator of ma- chine precision,

    Y. Wang and C.-Y. Lai “Multi-stage neural networks: Function approximator of ma- chine precision,”Journal of Computational Physics504(2024) 112865

    2024

  35. [35]

    Stacked networks im- prove physics-informed training: Applications to neural networks and deep operator networks,

    A. A. Howard, S. H. Murphy, S. E. Ahmed, and P. Stinis “Stacked networks im- prove physics-informed training: Applications to neural networks and deep operator networks,”Foundations of Data Science7(2025) no. 1, 134–162

    2025

  36. [36]

    warped-metrics

    S. L¨ ust, F. Ruehle, and S. Schreyer “warped-metrics.”https://github.com/ruehlef/ warped-metrics2026

  37. [37]

    Flux vacua of the mirror octic,

    E. Plauschinn and L. Schlechter “Flux vacua of the mirror octic,”JHEP01(2024) 157 [2310.06040]

    Pith/arXiv arXiv 2024

  38. [38]

    A Pair of Calabi-Yau manifolds as an exactly soluble superconformal theory,

    P. Candelas, X. C. De La Ossa, P. S. Green, and L. Parkes “A Pair of Calabi-Yau manifolds as an exactly soluble superconformal theory,”Nucl. Phys. B359(1991) 21– 74

    1991

  39. [39]

    Topological string theory on compact Calabi-Yau: Modularity and boundary conditions,

    M.-x. Huang, A. Klemm, and S. Quackenbush “Topological string theory on compact Calabi-Yau: Modularity and boundary conditions,”Lect. Notes Phys.757(2009) 45– 102 [hep-th/0612125]

    Pith/arXiv arXiv 2009

  40. [40]

    Mirror quintic vacua: hierarchies and inflation,

    N. Cabo Bizet, O. Loaiza-Brito, and I. Zavala “Mirror quintic vacua: hierarchies and inflation,”JHEP10(2016) 082 [1605.03974]

    Pith/arXiv arXiv 2016

  41. [41]

    Yukawa Couplings Between (2,1) Forms,

    P. Candelas “Yukawa Couplings Between (2,1) Forms,”Nucl. Phys. B298(1988) 458. 33

    1988

  42. [42]

    Moduli space of Calabi-Yau manifolds,

    P. Candelas and X. C. de la Ossa “Moduli space of Calabi-Yau manifolds,” inXIII International School of Theoretical Physics: The Standard Model and Beyond. 9, 1989

    1989

  43. [43]

    A landscape of orientifold vacua,

    F. Carta, J. Moritz, and A. Westphal “A landscape of orientifold vacua,”JHEP05 (2020) 107 [2003.04902]

    Pith/arXiv arXiv 2020

  44. [44]

    Level crossings, attractor points and complex multiplication,

    H. Ahmed and F. Ruehle “Level crossings, attractor points and complex multiplication,” JHEP06(2023) 164 [2304.00027]

    Pith/arXiv arXiv 2023

  45. [45]

    On the limited memory BFGS method for large scale opti- mization,

    D. C. Liu and J. Nocedal “On the limited memory BFGS method for large scale opti- mization,”Mathematical Programming45(1989) no. 1, 503–528

    1989

  46. [46]

    The Universal Kahler Modulus in Warped Compactifications,

    A. R. Frey, G. Torroba, B. Underwood, and M. R. Douglas “The Universal Kahler Modulus in Warped Compactifications,”JHEP01(2009) 036 [0810.5768]. 34

    Pith/arXiv arXiv 2009