GlobalCY I: A JAX Framework for Globally Defined and Symmetry-Aware Neural Kähler Potentials

This paper introduces GlobalCY, a JAX-based framework demonstrating that globally defined, symmetry-aware neural Kähler potential models outperform local-input baselines on challenging Calabi–Yau geometries by significantly reducing geometric inconsistencies like negative-eigenvalue frequency and projective-invariance drift.

The Gist

Imagine you are trying to teach a computer to understand the shape of a very complex, multi-dimensional object—a Calabi-Yau manifold. In the world of string theory (which tries to explain the universe's fundamental particles), these shapes are crucial. They are like the hidden gears inside a cosmic clockwork that determine how physics works.

The problem is that we can't write down a simple formula for the "perfect" shape of these gears. So, scientists use Neural Networks (AI) to learn and approximate these shapes.

However, there's a catch. The paper you shared, "GlobalCY," points out a major flaw in how these AI models are currently built. Here is the breakdown in simple terms:

The Problem: The "Local" vs. "Global" View

Think of the Calabi-Yau shape like a massive, intricate globe (like the Earth).

• The Old Way (Local-Input Models): Imagine you are trying to draw a map of the whole Earth, but you are only allowed to look at a tiny, 1-foot square of the ground at a time. You learn the texture of the grass and the rocks in that tiny square very well. But because you never see the big picture, your map might look great up close, but when you try to stitch all the squares together, the continents don't line up, the oceans are the wrong size, and the map falls apart.

  • In the paper: These models train well (low "loss") but fail when tested on the big picture. They break down near "singularities" (places where the shape gets twisted or sharp, like a mountain peak).

• The New Way (Global-Input Models): Now, imagine you are given a globe to look at. You can see how the continents curve and how the poles connect. You aren't just memorizing a patch of grass; you are learning the rules of how the whole sphere works.

  • In the paper: This is the GlobalCY framework. It forces the AI to learn using "global" rules (mathematical symmetries) that apply to the whole shape, not just a tiny piece.

The Experiment: The "Cefalú" Stress Test

The authors didn't just test this on easy shapes. They tested it on the "Cefalú family," which are the "hard mode" levels of these shapes. These are the twisted, near-broken versions where the geometry is very fragile.

They compared three types of AI architects:

1. The Local Architect: Only looks at tiny patches.
2. The Global Architect: Looks at the whole shape and respects its symmetry.
3. The Symmetry-Aware Architect: The Global Architect, but with a special "symmetry cheat sheet" that tells it exactly how the shape repeats itself.

The Results: What Happened?

They ran the models through a rigorous test (using a fixed set of random seeds, like running a race three times to make sure the winner is consistent).

• The Winner: The Global Architect (the one that respects the whole shape) won hands down.

  • It made fewer "mathematical errors" (negative eigenvalues).
  • It stayed consistent even when the shape was stretched or rotated (projective invariance).
  • It performed best on the hardest shapes (specifically the $\lambda = 0.75$ case).

• The Runner-Up: The Symmetry-Aware Architect did better than the Local one, but it was actually worse than the plain Global Architect in this specific test.

  • Why? The paper suggests that while knowing the symmetry is a good idea, the current way they implemented it was a bit "clunky." It's like giving the architect a cheat sheet, but the sheet is so complicated they get confused. It's a promising idea, but not quite ready to beat the simpler Global approach yet.

The Big Takeaway

The main message of the paper is this: Don't just let the AI memorize the data; teach it the rules of the geometry.

If you build an AI to understand complex shapes, you have to force it to understand the whole shape, not just the pieces. If you don't, the AI might look smart during training, but it will fail spectacularly when you ask it to do real physics work, especially near the "cracks" in the universe (singularities).

The Toolkit: GlobalCY

The authors didn't just find a better model; they built a toolkit called GlobalCY (written in a programming language called JAX).

• Think of this as a new construction site where you can easily swap out different types of architects (Local vs. Global) and see exactly who builds the best house, using the exact same blueprints and materials.
• This ensures that the results are fair and reproducible, so other scientists can trust the findings.

In a Nutshell

The paper says: "We built a better way to teach AI about the shape of the universe. By forcing the AI to look at the 'big picture' (global symmetry) instead of just 'close-ups' (local patches), we get much more accurate and stable results, especially when the shapes get weird and broken. This is a crucial step toward using AI to solve real problems in theoretical physics."

Technical Summary

1. Problem Statement

The central challenge addressed is the geometric fidelity of machine-learned Kähler potentials on Calabi–Yau (CY) manifolds. While neural networks can successfully minimize training loss (approximation error), they often fail to respect the underlying global geometric structure of the manifold.

• The Specific Issue: Local-input neural models (which learn corrections based on local coordinate charts) may appear successful during training but fail critical geometric diagnostics in "hard" regimes, particularly near singular or near-singular members of the Cefalú family of quartic hypersurfaces.
• The Failure Modes: These models often exhibit instability under projective rescaling, produce metrics with negative eigenvalues (violating the positive-definiteness required for a metric), and lack consistency across different coordinate charts.
• The Goal: To determine if imposing global invariant structure and symmetry-awareness as architectural constraints improves the behavior of learned Kähler potentials in these difficult regimes, rather than relying solely on optimization capacity.

2. Methodology

The paper introduces GlobalCY, a JAX-native framework designed to isolate architectural effects from data preprocessing or geometry generation artifacts.

A. Framework Architecture

The system is built on a two-layer stack:

1. GeoCYData: Provides the geometric substrate, including quartic geometry cases, bundle generation (local charts, invariant features, symmetry-aware views), and benchmark protocols.
2. GlobalCY: The learned-model layer that consumes these bundles to instantiate models, construct metrics, and run diagnostics.

• Key Design Choice: The framework uses JAX to ensure that the geometric construction ($g = g_{FS} + \partial\bar{\partial}\phi$), the neural model, and the diagnostic evaluation all exist within a single differentiable environment, ensuring reproducibility and eliminating post-hoc numerical coupling.

B. Model Families Compared

The study compares three distinct architectural approaches on the same geometric data:

1. LocalPhiMLP (Baseline): A standard Multi-Layer Perceptron (MLP) consuming local chart coordinates. It learns a scalar correction $\phi$ without explicit global constraints.
2. GlobalInvariantPhi: A model trained on projective-invariant feature representations. The input features are constructed to be compatible with the global projective structure of the hypersurface, forcing the learned correction to respect global geometry.
3. SymmetryAwareGlobalPhi: A model that uses global invariant features but incorporates symmetry-aware information (e.g., canonical representatives, orbit metadata) to further condition the architecture on the discrete symmetries of the manifold.

C. Benchmark Protocol

• Geometric Testbed: The Cefalú quartic family, specifically the hard cases $\lambda = 0.75$ and $\lambda = 1.0$. These regimes are known to be geometrically fragile and sensitive to singularities.
• Evaluation: A fixed multi-seed protocol (seeds 7, 11, 19) to ensure results are not artifacts of random initialization.
• Diagnostics: Instead of relying solely on training loss, the study uses a geometry-aware diagnostic suite:
  • Negative-eigenvalue frequency: Measures how often the learned metric violates positive-definiteness.
  • Projective-invariance drift: Measures the deviation of the learned metric under projective rescaling.
  • Minimum eigenvalue behavior: Tracks the lower tail of the spectrum.
  • Chart consistency: Checks stability across local coordinate patches.

3. Key Contributions

1. GlobalCY Framework: The release of a reproducible, JAX-native framework that cleanly separates geometry generation, model architecture, and diagnostic evaluation for Calabi–Yau metric learning.
2. Controlled Architectural Comparison: A rigorous benchmark comparing local, globally invariant, and symmetry-aware models on hard quartic regimes, isolating the impact of architectural constraints.
3. Geometry-First Diagnostics: A shift in evaluation metrics from pure optimization loss to geometric stability (negative eigenvalues, projective drift), demonstrating that "good loss" does not equal "good geometry."
4. Reproducible Results: A "freeze_results" workflow that generates stable, manuscript-ready artifacts (tables, figures, JSON) from the benchmark, ensuring scientific interpretability.

4. Results

The benchmark yielded clear conclusions regarding the efficacy of global structure:

• Global Invariant Model is Superior: The GlobalInvariantPhi model outperformed the local baseline on the two most critical metrics in both $\lambda = 0.75$ and $\lambda = 1.0$ cases:
  • Negative-eigenvalue frequency: Significantly reduced (e.g., at $\lambda=0.75$, dropped from 0.0885 to 0.0417).
  • Projective-invariance drift: Substantially lowered (e.g., at $\lambda=0.75$, dropped from $4.39 \times 10^{-8}$ to $1.44 \times 10^{-8}$).
  • Training Loss: Also achieved lower loss than the local baseline.
• Regime Difficulty: The gains were most pronounced at $\lambda = 0.75$. The $\lambda = 1.0$ case was harder, showing a smaller separation between global and local models, confirming it as a more difficult geometric regime.
• Symmetry-Aware Model Performance: The SymmetryAwareGlobalPhi model improved upon the local baseline regarding projective drift but did not surpass the plain GlobalInvariantPhi model. It exhibited higher variability across seeds and worse performance on negative-eigenvalue frequency. This suggests that while symmetry-awareness is promising, the current implementation is not yet the dominant architectural ingredient.
• Nuance: The local baseline occasionally retained a better minimum eigenvalue mean, indicating that global structure does not improve every metric simultaneously, but it dominates on the primary stability and fidelity metrics.

5. Significance and Future Directions

• Architectural Principle: The paper establishes that global invariant structure is not merely a cosmetic choice but a scientifically necessary constraint for learned Kähler potentials in hard regimes. It proves that "geometry-first" modeling yields more stable and faithful metrics than purely local regression.
• Foundation for Physics: By stabilizing the metric layer, this work enables downstream computations (curvature, characteristic forms, Yukawa couplings) that would otherwise inherit the instability of local-only models.
• Future Work: The authors identify clear next steps:
  • Refining symmetry-aware architectures to surpass the global invariant baseline.
  • Symbolic Distillation: Compressing the learned neural surrogates into interpretable analytic formulas.
  • Moduli Dependence: Training models that vary coherently across parameter families rather than case-by-case.
  • Singularity-Aware Learning: Developing models specifically for the most singular limits (like $\lambda=1.0$).

In summary, GlobalCY I demonstrates that embedding global geometric constraints directly into the neural architecture is essential for producing reliable, physically meaningful metrics on Calabi–Yau manifolds, particularly in the challenging regimes where standard machine learning approaches fail.