Lectures on the Calabi-Yau Landscape

Jiakang Bao, Yang‐Hui He, Edward Hirst, Stephen PietromonacoarXiv 2020 · DOI · arXiv

Abstract

In these lecture notes, we survey the landscape of Calabi-Yau threefolds and the application of machine learning to explore it. We begin with the compact portion of the landscape, focusing on complete intersection Calabi-Yau varieties (CICYs) and elliptic fibrations. Non-compact Calabi-Yau manifolds are also discussed as they manifest in Type II superstring theories, arising as representation varieties of quivers used to describe gauge theories. Given the vast amount of Calabi-Yau data, we examine how machine learning can be applied to algebraic geometry and the string landscape. These notes are intended for graduate students in mathematics and physics, based on lectures given at the 2019 PIMS Summer School on Algebraic Geometry in High-Energy Physics.

Introduction: The String Theory Challenge

In string theory, the universe is modeled as a 10-dimensional construct. However, our physical perception is limited to four dimensions. The remaining six dimensions are theorized to be "compactified" into infinitesimal, complex geometric shapes known as Calabi-Yau (CY) manifolds. The specific topology and geometry of these manifolds are not merely mathematical curiosities; they dictate the fundamental physics of our universe, including the types of particles and forces that emerge in four-dimensional spacetime.

The primary challenge facing researchers is the sheer scale of the "string landscape." There are hundreds of millions of known Calabi-Yau manifolds, and the number of possible vacuum states is estimated to be as high as $10^{500}$. Calculating the topological properties of even a single manifold—such as its Hodge numbers—using traditional algebraic geometry methods like Gröbner basis calculations is computationally expensive and time-consuming. This paper explores a paradigm shift: using machine learning (ML) to navigate this landscape by identifying patterns that bypass traditional computation.

Representing Geometry as Data

To apply machine learning to algebraic geometry, the geometric structures must first be translated into a format recognizable by algorithms. The authors focus on several key datasets:

  • Complete Intersection Calabi-Yau (CICY) Manifolds: These are represented as "configuration matrices." A CICY matrix can be viewed as a grid where the entries correspond to degrees of polynomials. Remarkably, a typical 12x15 CICY matrix can be treated as a low-resolution pixelated image, where the integer values represent different "colors."
  • Weighted Projective Space ($WP^4$): This dataset consists of 7,555 vectors that define hypersurfaces. These vectors serve as the input features for predictive models.
  • Elliptic Fibrations: A smaller subset of CICYs (643 manifolds) was used to test the limits of ML on sparse data. To compensate for the small sample size, the authors employed data augmentation—permuting rows and columns to create thousands of mathematically equivalent but visually distinct training samples.

Methodologies: From CNNs to SVMs

The research utilizes a variety of supervised learning architectures to probe the properties of these manifolds:

  1. Convolutional Neural Networks (CNNs): Given that configuration matrices resemble images, CNNs are used to detect spatial correlations within the geometric data.
  2. Multi-Layer Perceptrons (MLP) and Support Vector Machines (SVM): These are employed for both binary classification (e.g., determining if a Hodge number exceeds a certain threshold) and regression (predicting exact integer values).
  3. Rigorous Evaluation: To ensure the models are truly learning the underlying geometry rather than exploiting data imbalances, the authors utilize the Matthews Correlation Coefficient (MCC), providing a more robust measure of accuracy than simple percentage-based metrics.

Key Insights and Results

The application of ML to the Calabi-Yau landscape yielded several groundbreaking results:

  • High-Speed Accuracy: For $WP^4$ hypersurfaces, models predicted whether a Hodge number was greater than 50 with 93.25% accuracy. This prediction occurs in seconds on a standard laptop, whereas traditional mathematical software might take hours to reach the same conclusion.
  • Topological Prediction: For CICYs, the models achieved 97% accuracy in predicting specific topological invariants ($h^{1,1} > 5$).
  • The Limits of AI: Interestingly, the authors found that while ML excels at "tensor-like" geometric problems, it struggles with pure number theory, such as predicting prime numbers. This suggests that the success of ML in geometry is due to deep, non-random structural correlations within algebraic varieties that the neural networks are able to extract.

Discussion and Future Directions

While the results are promising, the study highlights the "Black Box" problem: while we know the AI is accurate, the specific mathematical reasons why these patterns exist often remain obscure. The AI provides the "what," but the theoretical "why" is still a subject for future mathematical proof.

Looking forward, the research points toward the even larger landscape of Calabi-Yau four-folds, which are central to F-theory and contain an estimated $10^{272,000}$ vacua. Furthermore, the use of Reinforcement Learning (RL) is proposed as a method to "hunt" through the landscape for specific manifolds that yield the exact particle physics of the Standard Model. By treating the search for our universe as a game or an optimization problem, AI may eventually identify the specific shape of the extra dimensions.

Summary updated

Abstract

In these lecture notes, we survey the landscape of Calabi-Yau threefolds, and the use of machine learning to explore it. We begin with the compact portion of the landscape, focusing in particular on complete intersection Calabi-Yau varieties (CICYs) and elliptic fibrations. Non-compact Calabi-Yau manifolds are manifest in Type II superstring theories, they arise as representation varieties of quivers, used to describe gauge theories in the bulk familiar four dimensions. Finally, given the huge amount of Calabi-Yau data, whether and how machine learning can be applied to algebraic geometry and string landscape is also discussed. These notes are directed to the beginning graduate student interested in mathematics and in physics, and are based on lectures given by the 2nd author at the 2019 PIMS Summer School on Algebraic Geometry in High-Energy Physics at the University of Saskatchewan.
BibTeX
@misc{bao2020lectures,
  author = {Jiakang Bao and Yang‐Hui He and Edward Hirst and Stephen Pietromonaco},
  title = {{Lectures on the Calabi-Yau Landscape}},
  year = {2020},
  eprint = {2001.01212},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.01212},
}
Figure