Metric Flows with Neural Networks

The Big Picture

Imagine trying to find the perfect shape of a multidimensional space. Not a sphere or a donut, but a complex geometric object called a Calabi-Yau manifold, which string theorists believe describes the hidden extra dimensions of our universe. Mathematicians proved these objects exist, but nobody has ever written down an explicit formula for one. It’s like knowing a perfect sculpture exists inside a block of marble, with no chisel and no instructions.

For decades, physicists and mathematicians have tried to approximate these elusive Riemannian metrics, the mathematical rulers that measure distances and angles on curved spaces. Classical methods like Donaldson’s algorithm can get close, but they’re painfully slow. Neural networks cracked that barrier: a modern laptop running the right network can match results that used to require far more computation. But why neural networks work so much better has remained mysterious.

James Halverson and Fabian Ruehle at Northeastern University and IAIFI have developed a mathematical theory explaining how neural networks flow through the space of geometric metrics. It unifies deep learning theory with the mathematics of curved spaces and reveals why neural networks beat classical approaches like Ricci flow, a technique that smooths out geometric irregularities the way heat spreads through a material.

Key Insight: Neural networks learn Calabi-Yau metrics better than fixed-kernel methods like Ricci flow precisely because they learn features. Their internal representation evolves during training in a way that fixed-kernel methods fundamentally cannot.

How It Works

To approximate a Calabi-Yau metric with a neural network, you encode the metric as the network’s output, then train it to satisfy Ricci-flatness, the requirement that the space has no net curvature anywhere. But what’s actually happening mathematically as gradient descent nudges the network’s weights toward a solution?

Halverson and Ruehle answered this by extending a powerful tool from machine learning theory: the Neural Tangent Kernel (NTK). The NTK describes how a network’s output changes in response to small parameter adjustments during training. For the metric learning setting, they derive a metric-NTK, a kernel governing how the metric at one point in space is linked to the metric at other points through the network’s shared parameters. This kernel behaves differently depending on the network’s width:

• Infinite-width limit: The metric-NTK freezes at initialization and never changes. The flow is mathematically clean but rigid, amounting to a linear kernel method.
• Finite-width networks: The metric-NTK evolves during training, coupling different regions of the manifold in a complex, time-dependent way.

With careful architectural choices (networks that couple metric values only to nearby points) and a specific loss function, they show that the infinite-width, fixed-kernel flow reproduces Perelman’s formulation of Ricci flow. Grigori Perelman reformulated Ricci flow as minimizing a single geometric quantity called the F-functional in his 2003 proof of the Poincaré conjecture, and the infinite-width neural network architecture recovers exactly this structure. Ricci flow, one of the most powerful tools in modern geometry, turns out to be a special case of neural network metric flow.

But those special conditions (infinite width, fixed kernel, local couplings) look highly restrictive from a machine learning perspective. The experiments confirm it.

The team tested fixed-kernel versus evolving-kernel networks on the quintic threefold, a classic Calabi-Yau manifold. Fixed-kernel networks, mimicking infinite-width Ricci-flow dynamics, learned the metric poorly; errors stayed high. Finite-width networks with an evolving metric-NTK converged to accurate metrics. Measuring the kernel at different training stages confirmed the picture: for well-trained finite-width networks, the kernel changed substantially from initialization to convergence. For fixed-kernel networks, it stayed frozen, and the results suffered.

Why It Matters

Accurate numerical Calabi-Yau metrics are needed to extract predictions from string theory compactifications: particle masses, coupling constants, features of the Standard Model. Better numerical methods bring string theory closer to making contact with experiment.

Mathematically, the work shows that Ricci flow is embedded inside a much richer family of flows parameterized by neural network architectures. This points toward a whole zoo of geometric flows with no classical names yet, generated by different network designs and loss functions. Some might converge faster, handle more complex geometries, or have properties that mathematicians would find genuinely useful.

The results also give machine learning theorists something concrete to chew on. The NTK framework makes falsifiable predictions here, and the experiments verify them. Feature learning (the ability of finite-width networks to evolve their internal representations) gives a quantifiable advantage over kernel methods, measured in geometric error rather than classification accuracy. Having a problem where you can pin down exactly why feature learning helps, with a geometric quantity you can compute, is valuable as neural networks move into more scientific domains.

So what comes next? Can different neural architectures realize other classical geometric flows? What are the fixed points for neural network metric flows in general? And can these methods scale to higher-dimensional Calabi-Yau manifolds with more complex topology?

Bottom Line: By building a mathematical theory of how neural networks flow through the space of geometric metrics, Halverson and Ruehle explain why neural networks outperform classical approaches like Ricci flow, and show that feature learning, not just optimization power, is the decisive advantage.