Neural and numerical methods for $\mathrm{G}_2$-structures on contact Calabi-Yau 7-manifolds

1. Problem Statement

The paper addresses the significant challenge of obtaining explicit metrics and geometric structures for $G_2$-manifolds.

• Context: $G_2$-manifolds (7-dimensional manifolds with holonomy group $G_2$) are crucial for M-theory compactifications, as they allow for minimal supersymmetry in 4D effective theories.
• The Obstacle: Unlike Calabi–Yau (CY) manifolds, which benefit from powerful algebraic geometry tools and extensive numerical methods for finding Ricci-flat metrics, $G_2$-manifolds lack such tools. Constructing explicit $G_2$-holonomy metrics requires solving a highly complex, nonlinear system of partial differential equations (PDEs) to find a torsion-free 3-form $\phi$ (satisfying $d\phi=0$ and $d*\phi=0$).
• Specific Gap: While topological invariants of $G_2$-manifolds are known, explicit geometric data (metrics, cycle volumes, Yukawa couplings) remains inaccessible. The authors aim to bridge this gap by using machine learning to approximate the defining 3-form and the induced metric on a specific class of $G_2$-structures: contact Calabi–Yau (cCY) manifolds.

2. Methodology

The authors propose a three-stage numerical and machine learning framework:

Stage 1: Approximating the Base Calabi–Yau Metric

• Base Manifold: The study focuses on the Fermat quintic threefold in $\mathbb{CP}^4$ as the base.
• Tool: They utilize the cymetric library (from prior work [12]), which employs neural networks to approximate Ricci-flat metrics on CY hypersurfaces.
• Process:
  • A neural network learns pointwise corrections to the Fubini–Study metric using a multiplicative Ansatz: $g_{pred} = g_{FS} + g_{FS} \odot g_{NN}$.
  • The model is trained on 200,000 points sampled from the quintic, minimizing a composite loss function that enforces the Monge–Ampère equation, Kählerity, and volume constraints.
  • Validation: The authors perform an independent check using a numerical exterior derivative (NED) algorithm to verify that the predicted Kähler form $\omega_{pred}$ is approximately closed ($d\omega \approx 0$), observing a scale-independent residual error of $O(10^{-1})$.

Stage 2: Constructing the $G_2$-Structure Data

• Geometric Setup: They construct the Calabi–Yau link ($K_f$), a 7-dimensional manifold defined as the intersection of the quintic hypersurface with the unit sphere $S^9$. This link is an $S^1$-bundle over the base CY threefold.
• Theoretical Basis: Using Proposition 7, they leverage the fact that a cCY manifold carries a coclosed $G_2$-structure (where $d*\phi=0$ but $d\phi \neq 0$).
  • The $G_2$ 3-form is defined as: $\phi = \eta \wedge \omega + \text{Re}(\Upsilon)$.
  • Here, $\eta$ is the contact 1-form, $\omega$ is the Kähler form (from Stage 1), and $\Upsilon$ is the holomorphic volume form.
• Data Generation:
  • Points on the 7-manifold are generated by sampling points on the base CY, lifting them to the link, and applying a phase rotation ($S^1$ action).
  • This yields a dataset of 900,000 points (split 90/5/5 for train/val/test).
  • At each point, the Kähler form and holomorphic volume form are computed (via residues for $\Upsilon$) and lifted to construct the full 3-form $\phi$, its Hodge dual $\psi$, and the induced metric $g_\phi$.
  • Consistency Checks: The dataset validates the fundamental $G_2$ identity $\phi \wedge \psi = 7 \text{vol}(g_\phi)$ with high precision.

Stage 3: Neural Learning of the $G_2$-Structure

• Architecture: A dense feed-forward neural network is designed to learn the mapping from geometric coordinates to the tensor components.
  • Input (19D): Ambient coordinates of the link point ($R^{10}$), local coordinates of the contact form $\eta$ ($R^7$), and patch indices (to handle coordinate charts).
  • Output:
    • Model A ($\phi$): Predicts the 35 independent components of the 3-form (antisymmetric $7\times7\times7$ tensor).
    • Model B ($g_\phi$): Predicts the 28 independent components of the metric (symmetric $7\times7$ tensor) via Cholesky decomposition ($g = LL^T$) to ensure positive definiteness.
  • Training: Uses Adam optimizer with Mean Squared Error (MSE) loss over 150 epochs. Normalization layers are applied to handle the varying scales of tensor components.

3. Key Contributions

1. First Numerical Approximation of $G_2$-Structures: This work presents the first successful application of neural networks to approximate the defining 3-form and induced metric of a $G_2$-structure on a contact Calabi–Yau link.
2. Hybrid Numerical-ML Pipeline: It establishes a robust pipeline combining symbolic/algebraic geometry (for constructing the link and volume forms) with deep learning (for metric approximation and interpolation).
3. Validation of Torsion via Neural Networks: The authors demonstrate how to use the trained neural network as a continuous function to compute numerical torsion ($d\phi$ and $d\psi$) via the numerical exterior derivative (NED). This validates that the learned structure satisfies the theoretical coclosed condition ($d\psi \approx 0$) and the specific torsion relation ($d\phi = \omega \wedge \omega$).
4. Statistical Analysis: The paper provides a detailed statistical analysis of the learned tensors, noting that while the 3-form components exhibit multimodal distributions (due to algebraic constraints), the metric components are smoother, making the metric a slightly easier regression target.

4. Results

• Accuracy:
  • 3-Form Model: Achieved a test MSE of $1.54 \times 10^{-6}$ with a Pearson correlation coefficient (PMCC) $\approx 1$.
  • Metric Model: Achieved a test MSE of $2.28 \times 10^{-6}$ with PMCC $\approx 1$.
• Torsion Verification:
  • Using the trained models to compute the exterior derivative, the authors found:
    • $\|d\psi\| = 0.074 \pm 0.036$ (confirming the coclosed condition $d*\phi=0$).
    • $\|d\phi\| / \|\omega \wedge \omega\| = 2.77 \pm 0.98$ (confirming the specific torsion relation for cCY manifolds).
  • The ratio $\phi \wedge \psi / \text{vol}(g_\phi)$ remained stable around the theoretical value of 7.
• Interpolation: The trained networks provide smooth interpolants, allowing for the evaluation of geometric quantities at points not present in the original training dataset.

5. Significance and Future Prospects

• Bridging Theory and Computation: This work demonstrates that machine learning can effectively handle the complex nonlinearities of exceptional holonomy geometries, opening the door to computing physical quantities (like cycle volumes and Yukawa couplings) in M-theory compactifications that were previously inaccessible.
• Foundation for Flows: The learned metrics serve as high-quality initial data for geometric flows (e.g., the Laplacian coflow). The authors propose using Physics-Informed Neural Networks (PINNs) in future work to evolve these coclosed structures toward fully torsion-free $G_2$-holonomy metrics.
• Scalability: The framework is not limited to the Fermat quintic; it can be extended to other weighted projective Calabi–Yau bases, enabling a systematic exploration of the landscape of $G_2$-manifolds.

In summary, the paper successfully establishes a "numerical geometry" toolkit for $G_2$-manifolds, proving that neural networks can learn and validate the intricate algebraic and differential structures required for M-theory compactifications.