# AI Meets String Theory: Mapping Calabi–Yau Spaces and the Quantum Landscape

- Link: https://jarlhalla.com/ai-meets-string-theory-mapping-calabi-yau-spaces-and-the-quantum-landscape/
- Published: 2026-08-08T08:53:34+02:00
- Author: knoksen

_A research-based professional article, updated 7 August 2026. Citation style: APA
7th edition._

## Abstract

String theory promises a quantum framework in which matter, forces, and gravity 
arise from a common microscopic structure. Its principal obstacle is no longer merely
writing consistent equations; it is determining which of an enormous number of mathematically
permitted solutions could describe our universe. Calabi–Yau manifolds sit at the
center of this problem. Their topology, complex structure, Kähler geometry, bundles,
branes, and fluxes help determine the fields and interactions of the resulting lower-
dimensional theory. Yet even sharply delimited construction classes contain hundreds
of millions of combinatorial starting points and vastly more triangulations, geometric
phases, bundles, and flux assignments. The resulting “landscape” cannot be exhaustively
searched by conventional symbolic and numerical methods.

Artificial intelligence is beginning to change the character of this research. Supervised
models classify topological invariants; graph neural networks learn triangulation
structure; reinforcement learning and genetic algorithms search sparse model-building
spaces; conditional generative models target rare flux vacua; and physics-informed
neural networks approximate Ricci-flat metrics and quantities closer to observable
particle physics, including normalized Yukawa couplings. This article traces the
historical path from dual-resonance models to Calabi–Yau compactification, explains
the landscape and swampland problems, evaluates representative AI applications, 
and develops a research agenda for trustworthy AI-assisted string phenomenology.
Its central argument is deliberately measured: AI has not selected the vacuum of
our universe or made string theory experimentally verified. It has, however, created
a new computational layer between abstract geometry and low-energy physics—one capable
of turning previously inaccessible calculations into testable, auditable, and sometimes
interpretable scientific workflows.

**Keywords:** string theory; Calabi–Yau manifold; compactification; string landscape;
swampland; machine learning; graph neural network; symbolic computation; flux vacuum;
string phenomenology

## 1. Introduction: From a Theory of Everything to a Search Problem

The most consequential questions in fundamental physics often appear in pairs. General
relativity explains gravity as the geometry of spacetime, while quantum field theory
explains particles and three nongravitational forces with extraordinary precision.
Each framework is successful within its domain, but their direct combination fails
at sufficiently high energies: treating Einstein gravity as an ordinary quantum 
field theory produces ultraviolet divergences requiring an unlimited tower of counterterms.
Near spacetime singularities, inside black holes, and around the Planck scale, the
separation between quantum matter and classical geometry becomes untenable.

String theory addresses this tension by replacing point particles with extended 
one-dimensional objects. Their vibrational modes include matter-like states, gauge
bosons, and—crucially—a massless spin-two excitation with the interactions expected
of a graviton. The extended nature of the string softens short-distance behavior,
while supersymmetry and higher-dimensional consistency tightly constrain the theory.
But this apparent unification comes with a price. Superstring theory is naturally
formulated in ten spacetime dimensions, so six spatial dimensions must be hidden
or otherwise rendered inaccessible at observed energies. The shape of those dimensions
is not inert scaffolding. It acts more like a microscopic instrument: its cycles,
curvature, singularities, bundles, branes, and fluxes determine which fields can
exist, their symmetries, and how they interact.

Calabi–Yau threefolds became the canonical six-real-dimensional compact spaces because
they can preserve a controlled amount of supersymmetry in four dimensions. They 
translate central physical questions into geometry. How many chiral families appear?
Which gauge group survives? Are there Higgs multiplets or fractionally charged exotics?
Can moduli be stabilized? What are the Yukawa couplings? In a compactification, 
these questions are encoded in topological invariants, cohomology groups, intersection
numbers, differential equations, and quantized fluxes.

The problem is scale. There are 7,890 standard complete-intersection Calabi–Yau (
CICY) configuration matrices, while the Kreuzer–Skarke classification contains 473,800,776
four-dimensional reflexive polytopes and 30,108 distinct Hodge-number pairs (Kreuzer&
Skarke, 2000). Each polytope may admit many fine, regular, star triangulations, 
and different triangulations may be redundant, birationally related, or physically
distinct after additional data are specified. Fluxes then multiply the possibilities.
The famous figure (10^{500}) is best understood as an order-of-magnitude emblem 
for a huge flux-vacuum discretuum, not as a census of Calabi–Yau manifolds. In a
particular F-theory estimate, one elliptically fibered fourfold was argued to support
on the order of (10^{272,000}) flux vacua (Taylor & Wang, 2015). These figures count
different objects under different assumptions, but they convey the same computational
reality: brute-force enumeration cannot be the sole strategy.

This is where AI becomes scientifically interesting. The relevant question is not
whether a neural network can imitate a table of known invariants. It is whether 
a hybrid system—combining exact geometry, symbolic algebra, numerical analysis, 
physical constraints, uncertainty quantification, and learned search—can discover
structure in a space too large for direct exploration. The purpose of this article
is to explain how that program emerged, what has already been achieved, where claims
must be qualified, and what would be required for AI-assisted landscape mapping 
to produce credible physical insight.

## 2. Historical Context: Why Geometry Entered Quantum Gravity

### 2.1 From hadronic amplitudes to quantum strings

The story began far from quantum gravity. Veneziano’s crossing-symmetric amplitude
was introduced in 1968 to model hadronic scattering and Regge behavior (Veneziano,
1968). Subsequent work revealed that the underlying mathematics described the vibrations
of a relativistic string. Quantum consistency required additional dimensions, and
the original bosonic theory contained both an unwanted tachyon and a massless spin-
two state. The latter, initially an inconvenience for a strong-interaction model,
became decisive: a consistent interacting massless spin-two particle reproduces 
gravity at low energy.

Worldsheet supersymmetry led to superstring theories without the bosonic tachyon
and with spacetime fermions. By the early 1980s, several formulations—Type I, Type
IIA, Type IIB, and later the two heterotic theories—were known. The 1984 Green–Schwarz
anomaly-cancellation mechanism showed that ten-dimensional superstrings could be
quantum mechanically consistent for special gauge groups (Green & Schwarz, 1984).
The heterotic (SO(32)) and (E_8\times E_8) strings, constructed by combining left-
and right-moving sectors in different ways, made contact with grand-unified model
building (Gross et al., 1985).

String theory therefore changed status. It was no longer simply a model of hadrons;
it was a candidate framework for quantum gravity and gauge unification. Yet a ten-
dimensional consistent theory was not a four-dimensional model of nature. Compactification
became the bridge.

### 2.2 Compactification and the demand for preserved supersymmetry

Kaluza and Klein had already shown that a compact extra dimension could make gravity
in higher dimensions resemble gravity plus electromagnetism in lower dimensions.
String compactification generalizes this logic. One assumes, at least approximately,
a spacetime of the form

[
\mathcal{M}_{10}\simeq \mathcal{M}_{3,1}\times X_6,]

where (\mathcal{M}_{3,1}) is four-dimensional spacetime and (X_6) is a compact six-
dimensional internal space. The Kaluza–Klein modes associated with variation along(
X_6) become heavy when the compactification scale is sufficiently small, leaving
a lower-energy four-dimensional effective theory.

The geometry cannot be arbitrary. If one seeks four-dimensional (\mathcal{N}=1) 
supersymmetry—the minimal amount compatible with chiral matter—then in the simplest
unwarped heterotic construction the internal manifold should admit a covariantly
constant spinor. A compact Kähler threefold with (SU(3)) holonomy supplies precisely
this structure. It has vanishing first Chern class and admits a Ricci-flat Kähler
metric. Such spaces are known as Calabi–Yau threefolds.

The name joins two mathematical achievements. Eugenio Calabi conjectured conditions
under which a compact Kähler manifold admits a Kähler metric with prescribed Ricci
curvature. Shing-Tung Yau proved the relevant existence and uniqueness theorem in
the 1970s (Yau, 1978). For a compact Kähler manifold with (c_1(X)=0), each Kähler
class contains a unique Ricci-flat Kähler metric. The theorem guarantees that the
desired metric exists, but it generally does not give a closed-form expression. 
That distinction—existence without explicit construction—later became a prime opening
for numerical and machine-learning methods.

### 2.3 The 1985 Calabi–Yau compactification breakthrough

Candelas, Horowitz, Strominger, and Witten (1985) made Calabi–Yau spaces central
to particle physics by showing how ten-dimensional (E_8\times E_8) heterotic string
theory could be compactified on a Calabi–Yau threefold while retaining four-dimensional(\
mathcal{N}=1) supersymmetry. In the standard embedding, part of the gauge connection
is identified with the spin connection. The visible (E_8) is then broken to (E_6),
and topological data determine the net chiral spectrum. In the simplest version,
the net number of generations is related to half the Euler characteristic, (N_{\
text{gen}}=|\chi(X)|/2), subject to the assumptions of that embedding.

This result established a geometry-to-physics dictionary. Hodge numbers (h^{1,1})
and (h^{2,1}) count classes of Kähler and complex-structure deformations in common
settings. The Euler characteristic of a Calabi–Yau threefold satisfies

[
\chi(X)=2\big(h^{1,1}-h^{2,1}\big).]

Vector-bundle cohomology controls matter multiplets. Triple intersection numbers
enter volumes, gauge kinetic functions, and couplings. Discrete symmetries and the
fundamental group can permit Wilson lines that break a grand-unified group to the
Standard Model gauge group. The low-energy theory is not determined by topology 
alone, but topology supplies the first layer of the calculation.

### 2.4 Classification, mirror symmetry, dualities, and the second superstring revolution

The late 1980s and early 1990s brought explicit databases and powerful structural
insights. The CICY program encoded complete intersections in products of projective
spaces as integer configuration matrices and produced 7,890 configurations. This
was an early example of “big data” in geometry: a machine-readable collection constructed
before contemporary data science existed.

Mirror symmetry then revealed that pairs of topologically distinct Calabi–Yau manifolds
could yield equivalent physics, exchanging complex-structure and Kähler data so 
that

[
h^{1,1}(X)=h^{2,1}(X^\vee),\qquad h^{2,1}(X)=h^{1,1}(X^\vee).]

The duality enabled difficult enumerative-geometry calculations and showed that 
a list of manifolds is not the same thing as a list of inequivalent physical theories(
Candelas et al., 1991). Geometric transitions connected apparently different compactifications.
D-branes expanded the available model-building ingredients and gave nonperturbative
meaning to gauge sectors and charged matter.

In 1995, Witten argued that the five consistent superstring theories are limits 
of a deeper network of dualities, with eleven-dimensional supergravity emerging 
from strongly coupled Type IIA theory (Witten, 1995). This “second superstring revolution”
reframed the multiplicity of formulations as complementary descriptions of an underlying
M-theory. Maldacena’s AdS/CFT proposal later supplied a nonperturbative definition
of string theory in certain asymptotically anti-de Sitter backgrounds and transformed
the study of quantum gravity (Maldacena, 1998).

These breakthroughs did not remove vacuum multiplicity. They made it more intelligible.
Distinct geometries could be dual, connected, or different phases of a larger moduli
space; fluxes could generate potentials; and branes could engineer gauge theories.
The question evolved from “Can string theory produce four-dimensional physics?” 
to “How is the physically relevant region selected from its solution space?”

## 3. The Quantum Landscape: What Is Being Counted?

### 3.1 Geometry is the beginning, not the vacuum

A Calabi–Yau manifold is not by itself a complete string vacuum. Depending on the
string formulation, a compactification may also require a holomorphic vector bundle
or coherent sheaf, D-brane configuration, orientifold action, quantized background
fluxes, and nonperturbative effects. The construction must satisfy anomaly cancellation,
flux quantization, tadpole cancellation, supersymmetry or controlled supersymmetry
breaking, and equations of motion. Scalar fields describing sizes and shapes—moduli—
must be stabilized or otherwise made compatible with observation.

The physical map is therefore layered:

| Mathematical or string datum | Representative low-energy consequence | 
| (h^{1,1}), (h^{2,1}), and other cohomology data | Numbers of moduli and candidate multiplets | 
| Intersection form and Chern classes | Volumes, kinetic terms, anomaly conditions, and couplings | 
| Fundamental group and discrete symmetries | Wilson-line breaking and selection rules | 
| Vector bundle, sheaf, or brane data | Gauge group, chiral matter, and Higgs/exotic content | 
| Quantized fluxes | Superpotential, moduli stabilization, vacuum energy, and supersymmetry breaking | 
| Ricci-flat metric and harmonic representatives | Normalized kinetic terms, masses, Yukawa couplings, and Kaluza–Klein spectra |

Two manifolds with identical Hodge numbers can have different intersection forms
and Chern data. The same manifold can support many bundles, brane configurations,
and flux choices. Conversely, dualities may identify descriptions that look different.
Any claimed landscape count must therefore specify its unit of counting: polytopes,
triangulations, diffeomorphism classes, effective field theories, flux assignments,
metastable vacua, or observationally distinct predictions.

### 3.2 The size hierarchy: from 7,890 to (10^{272,000})

The CICY list is large but manageable. The Kreuzer–Skarke dataset is qualitatively
different: 473,800,776 reflexive four-polytopes encode toric ambient spaces for 
Calabi–Yau hypersurfaces (Kreuzer & Skarke, 2000). These polytopes yield 30,108 
Hodge pairs, but neither polytopes nor triangulations correspond one-to-one with
topologically distinct threefolds. Fine, regular, star triangulations resolve the
ambient toric geometry; multiple triangulations can describe different Kähler phases
of the same threefold or redundant presentations.

Demirtas, McAllister, and Rios-Tascon (2020) proved an upper bound of roughly (10
^{428}) on topologically inequivalent Calabi–Yau hypersurfaces arising from the 
Kreuzer–Skarke list, even though the corresponding bound on triangulations is much
larger. A 2026 preprint further tightens the diffeomorphism-class upper bound to
approximately (10^{296}), illustrating how rapidly the combinatorial picture continues
to evolve (MacFadden et al., 2026). At low (h^{1,1}), exact counting is becoming
possible: Gendler et al. (2023) reported 4, 27, 183, 1,184, and 8,036 topological
equivalence classes for (h^{1,1}=1,2,3,4,) and (5), respectively, within the relevant
hypersurface construction.

Flux choices generate a different explosion. Bousso and Polchinski (2000) showed
how many quantized four-form fluxes can produce a dense discretuum of vacuum energies.
Flux compactifications later made estimates near (10^{500}) widely discussed, although
the result depends on the geometry, tadpole bounds, approximations, and the definition
of a vacuum. Taylor and Wang’s (2015) (\mathcal{O}(10^{272,000})) estimate concerns
flux choices on a particular F-theory fourfold and should not be conflated with 
the number of Calabi–Yau threefolds.

The lesson is not that one exponent has replaced another. It is that landscape size
is stratified. Geometry, resolution, bundle or brane data, fluxes, and moduli solutions
each introduce a combinatorial layer. Traditional exact methods remain indispensable
for certification, but they cannot be applied indiscriminately to every candidate.

### 3.3 Why finding Standard-Model-like physics is hard

A promising compactification must pass a sequence of increasingly expensive filters.
At minimum, a realistic model seeks the Standard Model gauge group (SU(3)\times 
SU(2)\times U(1)), three chiral generations, an acceptable Higgs sector, no unobserved
chiral exotics, viable anomaly cancellation, and mechanisms for neutrino masses 
and proton stability. It must then address moduli stabilization, supersymmetry breaking
or its absence at accessible energies, the cosmological constant, flavor hierarchies,
and cosmological history.

Topological filters are comparatively cheap; physical normalization is not. Cohomology
can count candidate zero modes, but physical masses and couplings require kinetic
terms. Those depend on the metric and harmonic representatives. Yau’s theorem does
not hand researchers a formula for that metric. Even before phenomenology, triangulations,
Gröbner bases, sheaf cohomology, period integrals, and nonlinear partial differential
equations can become computational bottlenecks.

Rare-target search makes the problem worse. A viable solution may occupy a tiny 
fraction of a high-dimensional discrete space. Uniform random sampling then wastes
most computation on obvious failures. Markov-chain methods can mix slowly between
disconnected or weakly connected regions. Hand-designed heuristics inherit human
biases toward familiar, low-Hodge-number, or computationally convenient geometries.

### 3.4 The swampland as a second classification problem

The landscape consists of low-energy effective theories believed to arise from consistent
quantum gravity. The “swampland” is the larger set of apparently consistent effective
field theories that cannot be completed into quantum gravity (Vafa, 2005). Swampland
conjectures attempt to identify structural boundaries without constructing every
ultraviolet completion.

Prominent examples include the Weak Gravity Conjecture, which in a basic form requires
a sufficiently light charged state so that extremal charged black holes can decay(
Arkani-Hamed et al., 2007); the Distance Conjecture, which predicts an infinite 
tower of increasingly light states at infinite distance in moduli space (Ooguri &
Vafa, 2007); and de Sitter conjectures that constrain positive scalar potentials(
Obied et al., 2018). These proposals are not equivalent in evidential status. The
Weak Gravity and Distance conjectures have substantial support across string constructions
and black-hole arguments but remain conjectures. Strong de Sitter statements are
particularly contested because controlled constructions and no-go claims depend 
sensitively on approximations, corrections, and stability criteria.

For AI, the swampland creates both an opportunity and a warning. A learned classifier
might identify empirical regularities separating known string constructions from
synthetic effective theories. But if the training set covers only a narrow construction
class, it may learn the conventions of the database rather than a universal principle
of quantum gravity. Swampland learning therefore demands out-of-distribution testing,
interpretable features, and analytic follow-up. Accuracy alone cannot turn a pattern
into a theorem.

## 4. How AI Maps Calabi–Yau Geometry

### 4.1 Machine learning for classification and regression

Many Calabi–Yau constructions have naturally tensorial encodings. A CICY is represented
by a sparse matrix of nonnegative integers describing multidegrees of defining polynomials.
A reflexive polytope can be represented by lattice points, vertices, faces, or invariant
coordinates. A triangulation can be encoded through simplices, height vectors, secondary-
cone data, or restrictions to two-faces. These are suitable inputs for neural networks,
support-vector machines, tree ensembles, and equivariant architectures.

The first wave of studies treated known datasets as controlled laboratories and 
articulated machine learning as a framework for “deep data dives” and conjecture
generation in the landscape (Carifio et al., 2017). If a model could infer Hodge
numbers, favorability, discrete symmetries, or cohomology from defining data, then
it might later prioritize expensive calculations in unknown regions. Erbin and Finotello(
2021) obtained 97% accuracy for (h^{1,1}) using 30% of a CICY dataset for training
and 99% using 70%; (h^{2,1}) remained substantially harder, reaching about 50% in
their setup. This asymmetry is scientifically informative: a network’s performance
reflects not only architecture but also how transparently a target invariant is 
encoded in the input representation.

More recent work has moved beyond Hodge numbers. He, Yao, and Yau (2026) used convolutional
architectures to predict divisibility invariants derived from triple intersection
data, reaching about 90% accuracy in standard cross-validation. These invariants
help distinguish topology more finely than a Hodge pair. The progression—from coarse
labels to intersection structure—is essential because phenomenology depends on richer
data than “how many holes” a manifold possesses.

The best use of such predictors is triage. A learned model evaluates millions of
inexpensive candidates; exact algorithms then verify the small subset likely to 
possess a desired property. The model does not replace the definition of a Hodge
number or an intersection calculation. It reallocates exact computation toward high-
value regions.

### 4.2 Graph methods: learning relations rather than flat arrays

Calabi–Yau data are relational. A polytope is a network of lattice points, faces,
and incidence relations. A triangulation is a collection of simplices glued along
shared faces. Toric divisors intersect according to combinatorial rules. Geometric
transitions connect different phases. Flattening these objects into fixed-size arrays
can destroy permutation symmetry and locality.

Graph methods preserve this structure. Nodes can represent lattice points, simplices,
divisors, or vacua; edges can encode incidence, adjacency, intersection, a bistellar
flip, or a geometric transition. Message-passing graph neural networks then aggregate
local information while sharing parameters across variable-sized objects. Equivariant
models can respect lattice or permutation symmetries, reducing the burden of learning
physically irrelevant coordinate choices.

A notable 2026 preprint introduced _dualGNN_, an autoregressive message-passing 
model operating on a generalized dual graph of triangulations, with signed circuits
as edge labels (MacFadden, 2026). The model sampled fine, regular triangulations
and generalized to unseen polygons. In string-theory applications it sampled Calabi–
Yau threefolds at (h^{1,1}=86) and produced results consistent with uniformity at(
h^{1,1}=128), well beyond earlier learned sampling regimes. Because the work was
still a preprint as of August 2026, its claims require continued independent validation,
but it demonstrates why graph representation is more than a fashionable architecture:
triangulation is intrinsically a graph-structured generative process.

Graph theory also supports non-neural analysis. Cole and Shiu (2019) used topological
data analysis and persistent homology to characterize structure in ensembles of 
flux vacua. Rather than reducing a point cloud to means and variances, persistent
homology tracks connected components, loops, and higher-dimensional features across
distance scales. This can reveal clustering, voids, or sampling biases—features 
relevant to whether a scan has found a physical concentration or merely followed
an algorithmic artifact.

### 4.3 Symbolic computation and neuro-symbolic acceleration

The geometry-to-physics pipeline is built on exact algebra. Researchers compute 
Stanley–Reisner ideals, intersection rings, Chern classes, line-bundle cohomology,
periods, and solutions of polynomial systems. Packages such as PALP, SageMath, Mathematica,
Macaulay2, Singular, cohomCalg, and CYTools encode decades of mathematical knowledge.
CYTools, for example, was designed to triangulate reflexive polytopes and compute
topological data even in high-Hodge-number regions of the Kreuzer–Skarke list (Demirtas
et al., 2022).

AI is most credible when coupled to this symbolic layer. One pattern is learned 
algorithm selection: the system predicts which exact strategy, monomial order, triangulation
move, or simplification is likely to minimize cost. Peifer, Stillman, and Halpern-
Leistner (2020) used reinforcement learning for S-pair selection in Buchberger’s
algorithm, outperforming standard selection strategies in some polynomial distributions.
Although this work was not limited to Calabi–Yau problems, Gröbner bases are central
to computational algebraic geometry, so the method is directly relevant.

Another pattern is formula discovery. Klaewer and Schlechter (2019) found that naïvely
predicting line-bundle cohomology was inadequate in general, but unsupervised learning
exposed piecewise-polynomial phases. The result could be converted into analytic
formulae rather than left as a black-box approximation. Similarly, decision trees
can generate human-readable conditions that geometers attempt to prove. This is 
the ideal neuro-symbolic loop: learn a pattern, express it symbolically, verify 
it exactly, and then add the verified result to the computational toolkit.

### 4.4 Data-driven search: from scanning to inverse design

Classification asks, “What property does this candidate have?” Landscape exploration
often needs the inverse question: “Which candidate will have the property we want?”
Genetic algorithms, reinforcement learning, Bayesian optimization, and conditional
generative models are suited to this inverse-design problem.

In genetic search, each candidate is encoded as a chromosome, a fitness function
scores physical desirability, and mutation and crossover explore nonlocal regions.
Cole, Schachner, and Shiu (2019) applied genetic algorithms to Type IIB flux vacua
on a symmetric six-torus and near a conifold region of a Calabi–Yau hypersurface.
Their algorithms located vacua with targeted properties more effectively than comparison
searches and exposed population structures that can themselves suggest hidden organization.

Reinforcement learning treats model building as a sequence of decisions. An agent
modifies bundle data, receives rewards for satisfying topological and physical constraints,
and learns a policy. Constantin et al. (2022) applied this approach to heterotic(
SO(10)) grand-unified models with monad bundles. On two selected Calabi–Yau manifolds,
trained policies produced phenomenologically promising states in nearly all episodes
and yielded hundreds of new candidate models. This does not mean that nearly every
string vacuum is realistic; it means that, after training in a carefully defined
environment, the policy learned to navigate a sparse constrained subspace far more
efficiently than uninformed sampling.

The newest step is conditional generation. Krippendorf and Liu (2026) trained conditional
variational autoencoders to generate Type IIB flux configurations with target values
of quantities such as the flux superpotential while incorporating tadpole constraints
in the objective. In conifold and symmetric-torus experiments, the method achieved
an approximately (10^3) speedup over Metropolis sampling in narrow target ranges
and generated distinct configurations beyond its training set. Because this study
was published in _Journal of High Energy Physics_, it represents a particularly 
significant transition from proof-of-concept classification to peer-reviewed inverse
landscape search. A complementary peer-reviewed study used a Bayesian flow network
for discrete flux vectors and a transformer for conditional sampling, further demonstrating
that modern generative architectures can search for Type IIB vacua with specified
properties (Walden & Larfors, 2026).

### 4.5 Numerical metrics: learning the shape, not merely the topology

Topology determines which fields may occur, but normalized physical quantities require
geometry. A Ricci-flat metric (g) on a Calabi–Yau manifold satisfies a nonlinear
system equivalent, in an appropriate formulation, to a complex Monge–Ampère equation.
Traditional numerical schemes based on balanced metrics, energy functionals, and
finite-dimensional approximations are powerful but computationally demanding.

Neural approaches parameterize a metric, Kähler potential, or related tensor field
and minimize a loss encoding the governing geometry. Loss terms can penalize violation
of the Monge–Ampère equation, non-Kähler behavior, disagreement across coordinate
patches, incorrect Kähler class, and nonzero Ricci curvature. Holomorphic or symmetry-
aware architectures build known structure into the model rather than asking a generic
network to rediscover it (Ashmore et al., 2020; Douglas et al., 2022; Jejjala et
al., 2022).

The _cymetric_ package extended these ideas to complete-intersection and Kreuzer–
Skarke geometries across Kähler and complex-structure moduli (Larfors et al., 2022).
The later _cymyc_ framework modeled tensor fields of arbitrary degree and moduli-
space geometry with a high-performance, geometry-aware machine-learning implementation(
Butbaia et al., 2025). Group-invariant models further improved data efficiency and
respected discrete symmetries through fundamental-domain projections (Hendi et al.,
2025).

These developments matter because the metric opens the door to Kaluza–Klein spectra,
harmonic forms, moduli-space metrics, curvature corrections, and normalized couplings.
The learned object is not valuable because it is a neural network. It is valuable
if residuals are controlled, coordinate transitions are consistent, known limits
are reproduced, and independent numerical methods agree.

### 4.6 What “identifying a stable vacuum” actually requires

The phrase _stable vacuum_ compresses several tests. A candidate should be a stationary
point of the relevant scalar potential; it should have no disallowed tachyonic directions;
its moduli should be stabilized or phenomenologically acceptable; and its decay 
lifetime should exceed the required physical timescale. The compactification must
also obey global constraints such as flux quantization, anomaly and tadpole cancellation.
Finally, the approximations used to derive the effective potential must remain controlled:
large volume, weak coupling, suppressed higher-derivative corrections, or another
justified regime.

AI currently addresses fragments of this problem. A classifier can predict whether
a restricted construction satisfies known consistency labels. An optimizer can locate
flux assignments near extrema. A conditional generator can target a small superpotential
or a tadpole-compatible region. A metric network can improve kinetic and mass calculations.
None of these outputs alone certifies a metastable four-dimensional vacuum. Certification
still requires solving the equations, evaluating the Hessian in the correct field-
space metric, checking global consistency, and estimating corrections and decay 
channels.

The most defensible workflow is therefore a cascade. Learned models first rank or
generate candidates; differentiable surrogates refine them; symbolic and numerical
solvers establish stationarity and spectra; and exact or bounded calculations test
consistency and control. In this sense, AI can _identify promising stable-vacuum
candidates_ and dramatically reduce search cost. Describing it as having already
found realistic stable vacua would exceed the evidence.

## 5. Case Studies: From Geometric Labels to Physical Quantities

### 5.1 CICY inference as a benchmark for scientific generalization

CICY matrices are ideal benchmarks because their invariants are known and their 
representation is compact. Early random train–test splits demonstrated high interpolation
accuracy, but realistic landscape use is harder: one wants to train on computationally
easy, low-complexity manifolds and predict high-complexity cases. Bull et al. (2019)
explicitly trained on lower (h^{1,1}) examples and tested on higher ones, showing
that limited “seeding” with difficult cases could substantially improve extrapolation.

This exposes a general rule for AI in mathematical physics. Random cross-validation
can overstate usefulness when near-duplicate structures populate both sets. A credible
benchmark should split by complexity, topology, construction family, or symmetry
orbit. It should test calibrated uncertainty and require the model to recognize 
when an input lies outside its competence.

### 5.2 F-theory gauge groups and interpretable decision rules

Wang and Zhang (2018) trained decision trees to infer non-Higgsable gauge groups
on divisors in four-dimensional F-theory bases from local triple-intersection information.
Depending on divisor class, they reported 85%–98% out-of-sample accuracy. More importantly,
the trees generated analytic rules. The authors proved a subset and used them to
construct local configurations with gauge groups including (SU(3)).

This case is a model of interpretable AI-assisted theory work. The machine did not
merely label geometries; it identified concise local conditions, and human analysis
converted some of those conditions into mathematical statements. The lasting scientific
output is therefore not the classifier’s accuracy but the new rule set and its domain
of validity.

### 5.3 Learning string Standard Models

Heterotic line-bundle constructions offer labeled datasets with direct phenomenological
criteria. Earlier systematic searches found hundreds of models with the exact supersymmetric
Standard Model matter spectrum after quotienting by discrete symmetries and introducing
Wilson lines (Anderson et al., 2012). Deen et al. (2022) showed that relatively 
small neural networks could distinguish consistent line-bundle models with the desired
gauge group and chiral asymmetry from random candidates. Autoencoders also separated
the classes without direct supervision. Predicting the number of Higgs multiplets—
a nontopological and more delicate property—was harder and benefited from engineered
features and larger models.

The result clarifies what “estimating a low-energy spectrum” means in current practice.
AI can often predict discrete spectrum-related labels from bundle and topology data.
It cannot yet take an arbitrary compactification and reliably output the fully normalized
Standard Model, its masses, mixings, and cosmology. Each additional step requires
more geometry and stronger validation.

### 5.4 Precision string phenomenology and Yukawa hierarchies

The most direct bridge toward measurable particle physics is the calculation of 
normalized couplings. Holomorphic Yukawa couplings alone are insufficient; normalization
depends on matter-field kinetic terms, which depend on the Calabi–Yau metric and
harmonic representatives. Recent work compared normalized Yukawa couplings obtained
through moduli-space geometry, period methods, and machine-learned Ricci-flat metrics,
finding excellent agreement in selected examples (Butbaia et al., 2024).

Berglund et al. (2025) then numerically computed physical Yukawa couplings for several
heterotic models in the standard embedding and found natural hierarchies. This is
a genuine advance: AI-assisted metric technology contributed to quantities structurally
related to quark and lepton flavor. It is not yet a unique prediction of observed
masses. The studies cover specific geometries and embeddings, and the full chain
from a stabilized vacuum through renormalization-group evolution to experiment remains
incomplete. Nevertheless, the conceptual barrier has shifted. A calculation once
blocked by the absence of explicit Ricci-flat metrics has become numerically approachable.

### 5.5 Generative triangulations and living datasets

Generative modeling is beginning to expand, rather than merely label, Calabi–Yau
datasets. Yip et al. (2025) introduced _CYTransformer_ to generate fine, regular,
star triangulations of four-dimensional reflexive polytopes, including polytopes
unseen during training. The model could retrain on its own verified output, motivating
the proposed AICY platform as a continuously expanding community resource. MacFadden’s(
2026) graph-based approach attacks a related problem with an architecture tailored
to dual-graph structure.

These preprints point toward “living atlases” in which models propose new triangulations,
exact software verifies regularity and computes invariants, deduplication identifies
equivalences, and the resulting certified examples improve the next model. The essential
word is _certified_. Unverified self-training can amplify subtle invalidity or sampling
bias; exact geometric checks must remain inside the loop.

## 6. Connecting the Landscape to Observation

The phrase “connect abstract constructions with measurable physics” can suggest 
a shorter bridge than actually exists. The full sequence is long:

[
\text{compactification data}\rightarrow \text{4D effective theory}\rightarrow \
text{vacuum selection and stabilization}\rightarrow \text{masses and couplings}\
rightarrow \text{cosmological history}\rightarrow \text{observables}.]

AI can accelerate multiple arrows, but no current system controls the entire chain.
A topological classifier may predict a gauge sector but not the vacuum expectation
values. A metric network may calculate kinetic normalization on a chosen geometry
but not explain why that geometry was selected. A flux generator may target a small
superpotential but rely on an approximate effective description whose corrections
must be checked.

Even so, landscape mapping can sharpen experimental questions. Compactifications
motivate axion-like particles, moduli, hidden photons, dark gauge sectors, cosmic
superstrings, supersymmetric spectra, and correlated coupling structures. Machine-
guided scans can determine whether a proposed signature is generic, rare, or incompatible
within a specified construction class. Conversely, null results can eliminate regions
of model space if the mapping from microscopic data to observables is controlled.

The appropriate standard is conditional prediction: _given_ a construction class,
consistency assumptions, moduli-stabilization scenario, and measure, what distributions
of observables follow? Such statements are weaker than a unique prediction but stronger
than unconstrained possibility. They are testable when assumptions and uncertainties
are explicit.

## 7. Future Implications: The Frontier of AI-Assisted String Theory

### 7.1 Physics-informed, equivariant, and generative architectures

Future models will increasingly encode exact symmetries and constraints. Permutation
invariance, lattice transformations, patch consistency, Kählerity, flux quantization,
and tadpole bounds should be architectural or algorithmic guarantees wherever possible.
A model that can never generate an invalid triangulation is preferable to one that
generates quickly and discards most outputs. A metric ansatz that is positive and
Kähler by construction is safer than an unconstrained tensor predictor.

Generative models will move from broad sampling to multi-objective inverse design.
A researcher may request a compactification with a target gauge group, three net
generations, specified discrete symmetry, stabilized moduli, controlled coupling,
and an axion decay constant within a chosen range. Conditional diffusion models,
normalizing flows, autoregressive graph networks, and transformer policies could
propose candidates, while differentiable or surrogate physics supplies gradients
and exact software certifies outputs.

### 7.2 Active learning and uncertainty-aware computation

Landscape labels are expensive. Active learning can choose the next exact calculation
to maximize information rather than filling a database uniformly. The system would
seek examples near decision boundaries, in poorly sampled topological classes, or
where models disagree. Multi-fidelity learning could combine cheap topological proxies,
medium-cost symbolic calculations, and expensive metric solutions.

Uncertainty must be operational. Calibrated predictive distributions should decide
whether a candidate is accepted, rejected, or escalated to exact computation. Ensembles,
conformal prediction, Bayesian approximations, and explicit out-of-distribution 
detectors can help, but mathematical datasets create special hazards: symmetry-related
duplicates, sharp phase boundaries, and rare equivalence classes can invalidate 
ordinary statistical assumptions.

### 7.3 Neuro-symbolic discovery and formal verification

The most scientifically productive AI may be a conjecture generator connected to
proof tools. Decision trees, sparse regression, symbolic transformers, and attribution
methods can extract candidate formulae from high-performing predictors. Computer
algebra can test them on large exact datasets; theorem provers can formalize tractable
subclaims; mathematicians can identify the missing conceptual argument.

Yang-Hui He (2024) argues that a hybrid of human expertise and AI will become integral
to theoretical discovery while not replacing theorists in the foreseeable future.
The Calabi–Yau program supports that view. Neural networks are strong at recognizing
patterns across combinatorial data, while physicists and geometers supply definitions,
invariants, equivalence relations, consistency constraints, and explanations. The
partnership is asymmetric but complementary.

### 7.4 Differentiable simulation, high-performance computing, and shared infrastructure

Automatic differentiation already enables neural PDE solvers and differentiable 
observables. GPU-accelerated packages such as JAX and specialized Calabi–Yau libraries
allow metric and curvature calculations to be embedded in optimization loops. Sparse
tensor accelerators, distributed computing, and improved exact-arithmetic kernels
will expand the range of tractable Hodge numbers and moduli dimensions.

Shared infrastructure is just as important as hardware. A trustworthy landscape 
platform needs versioned datasets, provenance for every invariant, canonical encodings,
equivalence and duplicate metadata, reproducible training splits, exact validators,
and benchmark tasks that test extrapolation. Living projects such as AICY are promising,
but governance must prevent a fast-growing archive from becoming a self-referential
training corpus with uncertain validity.

Quantum computing is sometimes proposed as a future accelerator, but its near-term
role should not be exaggerated. Landscape problems contain hard optimization, linear-
algebra, and sampling subroutines that may eventually benefit from quantum algorithms.
At present, fault-tolerant resources and proven end-to-end advantages are absent.
Classical geometric AI and high-performance symbolic computation remain the actionable
frontier.

### 7.5 Interpretability, validation, and the measure problem

Interpretability is not a cosmetic requirement in fundamental physics. A highly 
accurate classifier may exploit the ordering of matrix rows, padding conventions,
database-generation artifacts, or a hidden correlation with computational complexity.
Symmetry tests, counterfactual inputs, feature ablations, and representation changes
should therefore accompany headline accuracy.

Validation has at least four levels:

 1. **Mathematical validity:** Is the generated object actually a smooth Calabi–Yau
    construction, or does it satisfy the claimed topological property?
 2. **Physical consistency:** Are anomalies and tadpoles canceled, fluxes quantized,
    approximations controlled, and instabilities absent within the stated regime?
 3. **Statistical reliability:** Does the model generalize across construction families
    and complexity scales, and is its uncertainty calibrated?
 4. **Phenomenological relevance:** Does the effective theory survive present experimental
    and cosmological constraints after normalization and running?

There is also a problem AI cannot solve by scale alone: the measure problem. Counting
vacua does not tell us how likely they are. Cosmological dynamics, selection effects,
duality redundancies, and the choice of measure can radically alter conclusions.
A generative model trained on database frequency learns the sampling distribution
of that database, not necessarily a physically meaningful probability distribution.
Claims about “typical string predictions” must therefore separate raw counts, algorithmic
sampling weights, and proposed cosmological measures.

### 7.6 What experimental verification could look like

The most credible route to experiment is not for AI to announce the unique Calabi–
Yau shape of nature. It is to derive robust correlations shared by broad, well-defined
classes of consistent vacua. Examples might include constrained axion spectra, coupling
relations in hidden sectors, upper bounds on field ranges, characteristic towers
of states, or restricted patterns of supersymmetry breaking.

If a conjectured swampland constraint can be translated into a rigorous bound on
inflation, dark energy, or gauge sectors, improved cosmological and particle data
could test it. If precision metric calculations reveal recurring Yukawa textures
across a statistically controlled family, flavor measurements could constrain that
family. If AI finds that a proposed low-energy effective theory is consistently 
absent despite exhaustive, diversity-aware searches—and an interpretable obstruction
is extracted—the negative result could motivate a new quantum-gravity principle.

None of these outcomes is guaranteed. The crucial transition will be from model 
discovery to obstruction discovery: not merely finding examples that work, but identifying
why broad classes cannot work. Negative structure is often more universal and more
experimentally useful than another isolated vacuum.

## 8. A Research Agenda for Trustworthy Landscape Mapping

The next stage of the field would benefit from a coordinated program with six priorities.

First, datasets should record provenance and equivalence. Every label should identify
the exact algorithm, software version, assumptions, and certification status that
produced it. Symmetry-related and birationally related examples should not leak 
across benchmark splits without explicit intent.

Second, benchmarks should emphasize extrapolation. Training on low (h^{1,1}) and
testing on high (h^{1,1}), transferring between CICY and toric constructions, or
testing on unseen symmetry classes is more informative than random splitting.

Third, generative systems should incorporate hard constraints. Exact validity masks,
equivariant representations, symbolic post-checks, and rejection reasons should 
be part of the model, not an afterthought.

Fourth, learned predictions should terminate in auditable physics. Candidate compactifications
must be followed through cohomology, moduli stabilization, metric approximation,
spectra, couplings, and uncertainty propagation. A topological hit is the start 
of analysis, not its conclusion.

Fifth, interpretable discoveries should be prioritized. New formulas, provable decision
rules, geometric obstructions, and certified sampling algorithms create reusable
knowledge. A black box with one percentage point more benchmark accuracy may not.

Sixth, interdisciplinary teams should be designed around the full chain. Algebraic
and differential geometers understand equivalence and invariants; string theorists
define consistency and physical targets; computer-algebra researchers build exact
engines; machine-learning researchers design representations and uncertainty methods;
phenomenologists and experimentalists identify which outputs could matter. Without
this collaboration, optimization risks becoming detached from either mathematics
or observation.

## 9. Conclusion

String theory’s landscape problem emerged from success. Quantum consistency, supersymmetry,
duality, branes, fluxes, and Calabi–Yau geometry produced an extraordinarily rich
set of possible lower-dimensional worlds. The same richness made vacuum selection
and systematic phenomenology computationally formidable. Hundreds of millions of
reflexive polytopes, vast triangulation spaces, bundle and brane choices, and astronomical
flux ensembles cannot be handled by naïve enumeration.

AI changes the available strategy. Classification models can infer topological and
spectrum-related properties; graph methods preserve the relational structure of 
triangulations and transitions; neuro-symbolic systems can accelerate exact algebra
and expose candidate formulae; reinforcement learning and evolutionary search can
navigate sparse model-building spaces; conditional generators can solve targeted
inverse problems; and physics-informed neural networks can approximate Ricci-flat
metrics and normalized couplings. The trajectory from Hodge-number prediction to
precision Yukawa calculations is particularly significant because it moves AI from
cataloguing geometry toward evaluating physics.

The transformation should not be overstated. No AI system has identified the vacuum
of our universe. The (10^{500}) landscape is not a single enumerated dataset. Swampland
conjectures remain conjectural, de Sitter constructions remain debated, and high
validation accuracy does not establish a theorem or a physical law. Database bias,
equivalence, uncertainty, stability, and the measure problem remain fundamental 
obstacles.

The responsible conclusion is nevertheless ambitious. AI can become a new scientific
instrument for string theory—not an oracle, but a mapper, accelerator, conjecture
engine, and inverse-design partner embedded in exact mathematics. Its most valuable
products may be certified geometries, interpretable constraints, reliable surrogates
for expensive calculations, and obstructions that reveal previously hidden principles
of quantum gravity.

The frontier now requires collaboration across geometry, physics, computer algebra,
machine learning, and phenomenology. If these communities build shared datasets,
symmetry-aware models, exact validation loops, and end-to-end uncertainty accounting,
the quantum landscape can become less a metaphor for intractable multiplicity and
more a structured domain of scientific inference. Whether that effort ultimately
singles out our universe, rules out broad classes of models, or uncovers new mathematics,
it will deepen our understanding of the relationship between information, geometry,
and physical law.

## References

Anderson, L. B., Gray, J., Lukas, A., & Palti, E. (2012). Heterotic line bundle 
standard models. _Journal of High Energy Physics, 2012_(6), 113. [https://doi.org/10.1007/JHEP06(2012)113](https://doi.org/10.1007/JHEP06(2012)113)

Arkani-Hamed, N., Motl, L., Nicolis, A., & Vafa, C. (2007). The string landscape,
black holes and gravity as the weakest force. _Journal of High Energy Physics, 2007_(
6), 060. [https://doi.org/10.1088/1126-6708/2007/06/060](https://doi.org/10.1088/1126-6708/2007/06/060)

Ashmore, A., He, Y.-H., & Ovrut, B. A. (2020). Machine learning Calabi–Yau metrics.
_Fortschritte der Physik, 68_(9), 2000068. [https://doi.org/10.1002/prop.202000068](https://doi.org/10.1002/prop.202000068)

Berglund, P., Butbaia, G., Hübsch, T., Jejjala, V., Mayorga Peña, D., Mishra, C.,&
Tan, J. (2025). Precision string phenomenology. _Physical Review D, 111_, 086007.
[https://doi.org/10.1103/PhysRevD.111.086007](https://doi.org/10.1103/PhysRevD.111.086007)

Bousso, R., & Polchinski, J. (2000). Quantization of four-form fluxes and dynamical
neutralization of the cosmological constant. _Journal of High Energy Physics, 2000_(
6), 006. [https://doi.org/10.1088/1126-6708/2000/06/006](https://doi.org/10.1088/1126-6708/2000/06/006)

Bull, K., He, Y.-H., Jejjala, V., & Mishra, C. (2019). Getting CICY high. _Physics
Letters B, 795_, 700–706. [https://doi.org/10.1016/j.physletb.2019.06.067](https://doi.org/10.1016/j.physletb.2019.06.067)

Butbaia, G., Mayorga Peña, D., Tan, J., Berglund, P., Hübsch, T., Jejjala, V., &
Mishra, C. (2024). Physical Yukawa couplings in heterotic string compactifications.
_Advances in Theoretical and Mathematical Physics, 28_(8), 2783–2822. [https://doi.org/10.4310/ATMP.241119041341](https://doi.org/10.4310/ATMP.241119041341)

Butbaia, G., Mayorga Peña, D., Tan, J., Berglund, P., Hübsch, T., Jejjala, V., &
Mishra, C. (2025). cymyc: Calabi–Yau metrics, Yukawas, and curvature. _Journal of
High Energy Physics, 2025_(3), 028. [https://doi.org/10.1007/JHEP03(2025)028](https://doi.org/10.1007/JHEP03(2025)028)

Candelas, P., de la Ossa, X. C., Green, P. S., & Parkes, L. (1991). A pair of Calabi–
Yau manifolds as an exactly soluble superconformal theory. _Nuclear Physics B, 359_,
21–74. [https://doi.org/10.1016/0550-3213(91)90292-6](https://doi.org/10.1016/0550-3213(91)90292-6)

Candelas, P., Horowitz, G. T., Strominger, A., & Witten, E. (1985). Vacuum configurations
for superstrings. _Nuclear Physics B, 258_, 46–74. [https://doi.org/10.1016/0550-3213(85)90602-9](https://doi.org/10.1016/0550-3213(85)90602-9)

Carifio, J., Halverson, J., Krioukov, D., & Nelson, B. D. (2017). Machine learning
in the string landscape. _Journal of High Energy Physics, 2017_(9), 157. [https://doi.org/10.1007/JHEP09(2017)157](https://doi.org/10.1007/JHEP09(2017)157)

Cole, A., Schachner, A., & Shiu, G. (2019). Searching the landscape of flux vacua
with genetic algorithms. _Journal of High Energy Physics, 2019_(11), 045. [https://doi.org/10.1007/JHEP11(2019)045](https://doi.org/10.1007/JHEP11(2019)045)

Cole, A., & Shiu, G. (2019). Topological data analysis for the string landscape.
_Journal of High Energy Physics, 2019_(3), 054. [https://doi.org/10.1007/JHEP03(2019)054](https://doi.org/10.1007/JHEP03(2019)054)

Constantin, A., Harvey, T. R., & Lukas, A. (2022). Heterotic string model building
with monad bundles and reinforcement learning. _Fortschritte der Physik, 70_(2–3),
2100186. [https://doi.org/10.1002/prop.202100186](https://doi.org/10.1002/prop.202100186)

Deen, R., Ovrut, B. A., Purves, A., & Sethi, S. (2022). Machine learning string 
standard models. _Physical Review D, 105_, 046001. [https://doi.org/10.1103/PhysRevD.105.046001](https://doi.org/10.1103/PhysRevD.105.046001)

Demirtas, M., McAllister, L., & Rios-Tascon, A. (2020). Bounding the Kreuzer–Skarke
landscape. _Fortschritte der Physik, 68_(11), 2000086. [https://doi.org/10.1002/prop.202000086](https://doi.org/10.1002/prop.202000086)

Demirtas, M., Rios-Tascon, A., & McAllister, L. (2022). CYTools: A software package
for analyzing Calabi–Yau manifolds. _arXiv_. [https://arxiv.org/abs/2211.03823](https://arxiv.org/abs/2211.03823)

Douglas, M. R., Lakshminarasimhan, S., & Qi, Y. (2022). Numerical Calabi–Yau metrics
from holomorphic networks. _Proceedings of Machine Learning Research, 145_, 223–
252. [https://proceedings.mlr.press/v145/douglas22a.html](https://proceedings.mlr.press/v145/douglas22a.html)

Erbin, H., & Finotello, R. (2021). Machine learning for complete intersection Calabi–
Yau manifolds: A methodological study. _Physical Review D, 103_, 126014. [https://doi.org/10.1103/PhysRevD.103.126014](https://doi.org/10.1103/PhysRevD.103.126014)

Gendler, N., MacFadden, N., McAllister, L., Moritz, J., Nally, R., Schachner, A.,&
Stillman, M. (2023). Counting Calabi–Yau threefolds. _arXiv_. [https://arxiv.org/abs/2310.06820](https://arxiv.org/abs/2310.06820)

Green, M. B., & Schwarz, J. H. (1984). Anomaly cancellations in supersymmetric (
D=10) gauge theory and superstring theory. _Physics Letters B, 149_, 117–122. [https://doi.org/10.1016/0370-2693(84)91565-X](https://doi.org/10.1016/0370-2693(84)91565-X)

Gross, D. J., Harvey, J. A., Martinec, E., & Rohm, R. (1985). Heterotic string. 
_Physical Review Letters, 54_, 502–505. [https://doi.org/10.1103/PhysRevLett.54.502](https://doi.org/10.1103/PhysRevLett.54.502)

He, Y.-H. (2024). AI-driven research in pure mathematics and theoretical physics.
_Nature Reviews Physics, 6_, 546–553. [https://doi.org/10.1038/s42254-024-00740-1](https://doi.org/10.1038/s42254-024-00740-1)

He, Y.-H., Yao, Z.-G., & Yau, S.-T. (2026). Distinguishing Calabi–Yau topology using
machine learning. _Harvard Data Science Review, 8_(1). [https://hdsr.mitpress.mit.edu/pub/gp2772c2](https://hdsr.mitpress.mit.edu/pub/gp2772c2)

Hendi, Y., Larfors, M., & Walden, M. (2025). Learning group invariant Calabi–Yau
metrics by fundamental domain projections. _Machine Learning: Science and Technology,
6_, 015050. [https://doi.org/10.1088/2632-2153/adb4bb](https://doi.org/10.1088/2632-2153/adb4bb)

Jejjala, V., Mayorga Peña, D. K., & Mishra, C. (2022). Neural network approximations
for Calabi–Yau metrics. _Journal of High Energy Physics, 2022_(8), 105. [https://doi.org/10.1007/JHEP08(2022)105](https://doi.org/10.1007/JHEP08(2022)105)

Klaewer, D., & Schlechter, L. (2019). Machine learning line bundle cohomologies 
of hypersurfaces in toric varieties. _Physics Letters B, 789_, 438–443. [https://doi.org/10.1016/j.physletb.2019.01.002](https://doi.org/10.1016/j.physletb.2019.01.002)

Kreuzer, M., & Skarke, H. (2000). Complete classification of reflexive polyhedra
in four dimensions. _Advances in Theoretical and Mathematical Physics, 4_, 1209–
1230. [https://arxiv.org/abs/hep-th/0002240](https://arxiv.org/abs/hep-th/0002240)

Krippendorf, S., & Liu, Z. (2026). Solving inverse problems of Type IIB flux vacua
with conditional generative models. _Journal of High Energy Physics, 2026_(7), 103.
[https://doi.org/10.1007/JHEP07(2026)103](https://doi.org/10.1007/JHEP07(2026)103)

Larfors, M., Lukas, A., Ruehle, F., & Schneider, R. (2022). Numerical metrics for
complete intersection and Kreuzer–Skarke Calabi–Yau manifolds. _Machine Learning:
Science and Technology, 3_, 035014. [https://doi.org/10.1088/2632-2153/ac8e4e](https://doi.org/10.1088/2632-2153/ac8e4e)

MacFadden, N. (2026). Sampling triangulations and Calabi–Yau threefolds with autoregressive
GNNs [Preprint]. _arXiv_. [https://arxiv.org/abs/2605.27770](https://arxiv.org/abs/2605.27770)

MacFadden, N., Orevkov, S. Y., & Stepniczka, M. (2026). Further bounding the Kreuzer–
Skarke landscape [Preprint]. _arXiv_. [https://arxiv.org/abs/2602.16909](https://arxiv.org/abs/2602.16909)

Maldacena, J. (1998). The large (N) limit of superconformal field theories and supergravity.
_Advances in Theoretical and Mathematical Physics, 2_, 231–252. [https://doi.org/10.4310/ATMP.1998.v2.n2.a1](https://doi.org/10.4310/ATMP.1998.v2.n2.a1)

Obied, G., Ooguri, H., Spodyneiko, L., & Vafa, C. (2018). De Sitter space and the
swampland [Preprint]. _arXiv_. [https://arxiv.org/abs/1806.08362](https://arxiv.org/abs/1806.08362)

Ooguri, H., & Vafa, C. (2007). On the geometry of the string landscape and the swampland.
_Nuclear Physics B, 766_, 21–33. [https://doi.org/10.1016/j.nuclphysb.2006.10.033](https://doi.org/10.1016/j.nuclphysb.2006.10.033)

Peifer, D., Stillman, M., & Halpern-Leistner, D. (2020). Learning selection strategies
in Buchberger’s algorithm. _Proceedings of Machine Learning Research, 119_, 7575–
7585. [https://proceedings.mlr.press/v119/peifer20a.html](https://proceedings.mlr.press/v119/peifer20a.html)

Taylor, W., & Wang, Y.-N. (2015). The F-theory geometry with most flux vacua. _Journal
of High Energy Physics, 2015_(12), 164. [https://doi.org/10.1007/JHEP12(2015)164](https://doi.org/10.1007/JHEP12(2015)164)

Vafa, C. (2005). The string landscape and the swampland [Preprint]. _arXiv_. [https://arxiv.org/abs/hep-th/0509212](https://arxiv.org/abs/hep-th/0509212)

Veneziano, G. (1968). Construction of a crossing-symmetric, Regge-behaved amplitude
for linearly rising trajectories. _Il Nuovo Cimento A, 57_, 190–197. [https://doi.org/10.1007/BF02824451](https://doi.org/10.1007/BF02824451)

Walden, M., & Larfors, M. (2026). Sampling string vacua using generative models.
_Machine Learning: Science and Technology, 7_, 015018. [https://doi.org/10.1088/2632-2153/ae32dc](https://doi.org/10.1088/2632-2153/ae32dc)

Wang, Y.-N., & Zhang, Z. (2018). Learning non-Higgsable gauge groups in 4D F-theory.
_Journal of High Energy Physics, 2018_(8), 009. [https://doi.org/10.1007/JHEP08(2018)009](https://doi.org/10.1007/JHEP08(2018)009)

Witten, E. (1995). String theory dynamics in various dimensions. _Nuclear Physics
B, 443_, 85–126. [https://doi.org/10.1016/0550-3213(95)00158-O](https://doi.org/10.1016/0550-3213(95)00158-O)

Yau, S.-T. (1978). On the Ricci curvature of a compact Kähler manifold and the complex
Monge–Ampère equation, I. _Communications on Pure and Applied Mathematics, 31_, 
339–411. [https://doi.org/10.1002/cpa.3160310304](https://doi.org/10.1002/cpa.3160310304)

Yip, J. H. T., Arnal, C., Charton, F., & Shiu, G. (2025). Transforming Calabi–Yau
constructions: Generating new Calabi–Yau manifolds with transformers [Preprint].
_arXiv_. [https://arxiv.org/abs/2507.03732](https://arxiv.org/abs/2507.03732)

AI Meets String Theory: Mapping Calabi–Yau Spaces and the Quantum Landscape