Machine Learning Algebraic Geometry for Physics

Jiakang Bao, Yang‐Hui He, Elli Heyes, Edward HirstarXiv 2022 · DOI · arXiv

This review paper, "Machine Learning Algebraic Geometry for Physics," provides a comprehensive look at how the "Big Data" revolution is meeting the abstract world of pure mathematics and string theory. The authors demonstrate that by treating geometric objects as tensors, machine learning (ML) can not only predict complex topological properties with high accuracy but also help mathematicians uncover new analytical insights.

TL;DR

This paper reviews how machine learning—ranging from simple regression to Siamese networks—is revolutionizing algebraic geometry and string theory. By treating geometric objects as tensors, researchers can predict complex topological properties and uncover hidden mathematical structures that were previously computationally prohibitive to explore.

What Problem Does This Paper Address?

The central challenge lies in the "String Landscape." String theory suggests our universe has 10 dimensions, meaning six must be "curled up" into a Calabi-Yau (CY) manifold. The problem? There are billions of candidate manifolds, and calculating their physical properties (like the number of particle generations) using traditional algebraic geometry is computationally exhausting.

The authors explore whether AI can "learn" the underlying rules of mathematics. Instead of solving grueling equations for every new manifold, can a neural network look at the "DNA" of a geometric shape (its weights or coordinates) and immediately tell us its topological features?

Methods

The paper surveys a diverse toolkit of ML techniques applied to different geometric problems:

  • Supervised Learning: Multi-layer Perceptrons (MLP) and Convolutional Neural Networks (CNN) are used for regression (predicting continuous values like volumes) and classification (identifying symmetry classes).
  • Unsupervised Learning: Principal Component Analysis (PCA) and K-Means clustering are used to find hidden patterns in high-dimensional data.
  • Topological Data Analysis (TDA): Specifically "Persistent Homology," which helps visualize the "shape" of the data itself to see if candidate manifolds naturally cluster together.
  • Siamese Neural Networks (SNN): Used to determine if two different-looking mathematical objects are actually "equivalent" under certain transformations.
  • Manifold Learning: Techniques like Multi-Dimensional Scaling (MDS) are used to project complex data into 2D/3D for human intuition.

Datasets

The authors utilize several landmark datasets in the field:

  • Kreuzer-Skarke Database: A massive collection of Calabi-Yau hypersurfaces.
  • Toric Fano Polytopes: 2D polygons and 3D polyhedra used in toric geometry.
  • Graded Ring Database (GRDB): Used for Hilbert Series data associated with algebraic varieties.
  • Brane Web Matrices: Data representing configurations of 5-branes in string theory.
  • Dessins d’Enfants: Bipartite graphs on Riemann surfaces used to study the absolute Galois group.

Key Insights and Ideas

  • Geometry as Tensors: The "aha" moment is that most problems in algebraic geometry can be reformulated as mappings between tensors. This makes them perfectly suited for standard deep learning architectures.
  • Representation is Everything: For "Dessins d’Enfants," the authors found that a standard adjacency matrix performed poorly. However, switching to a "cyclic edge list"—a representation that more faithfully captures the graph's embedding—boosted accuracy from near-random to over 90%.
  • Symbolic Discovery: In the study of "Amoebae" (projections of complex curves), the authors used the weights of a trained neural network to derive a simple, previously unknown analytical formula for the genus of the shape. This shows ML acting as a "collaborator" in mathematical discovery.
  • The GAN-Geometry Connection: The paper highlights a deep theoretical "dictionary" between the Monge-Ampère equations in K¨ahler geometry and the loss functions used in Generative Adversarial Networks (GANs) via the theory of Optimal Transport.

Key Results

  • Topological Prediction: Neural networks predicted Hodge numbers and Euler characteristics of CY manifolds with $R^2$ scores exceeding 0.94, suggesting the network "learned" the complex index theorems involved.
  • Volume Calculation: For 2D Fano polygons, MLPs achieved 99%+ accuracy in predicting normalized volumes.
  • Equivalence Testing: Siamese networks reached 100% accuracy in identifying "weak equivalence" in brane webs, though they struggled with "strong equivalence" (50% accuracy), indicating that some mathematical symmetries are still too subtle for current architectures.
  • Mutation Classes: Naive Bayes classifiers achieved 100% accuracy in distinguishing between different mutation classes of quivers (graphs representing gauge theories).

Limitations and Future Work

  • The "Strong Equivalence" Gap: The failure of ML to distinguish strong equivalence in brane webs suggests that standard NNs might miss deep, non-local mathematical properties. Future work may require "Physics-Informed" or "Symmetry-Preserving" architectures.
  • Data Quality: The authors noted that NNs could easily distinguish "real" mathematical data from "fake" synthetic data, highlighting the difficulty of generating representative samples for training.
  • AI-Assisted Conjectures: The ultimate goal is moving beyond mere prediction toward "conjecture formulation," where AI points humans toward new theorems by identifying patterns in data that are invisible to the naked eye.

Summary updated

Abstract

We review some recent applications of machine learning to algebraic geometry and physics. Since problems in algebraic geometry can typically be reformulated as mappings between tensors, this makes them particularly amenable to supervised learning. Additionally, unsupervised methods can provide insight into the structure of such geometrical data. At the heart of this programme is the question of how geometry can be machine learned, and indeed how AI helps one to do mathematics. This is a chapter contribution to the book Machine learning and Algebraic Geometry, edited by A. Kasprzyk et al.
BibTeX
@misc{bao2022machine,
  author = {Jiakang Bao and Yang‐Hui He and Elli Heyes and Edward Hirst},
  title = {{Machine Learning Algebraic Geometry for Physics}},
  year = {2022},
  eprint = {2204.10334},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2204.10334},
}
Figure