Characterizing 4-string contact interaction using machine learning

The Big Picture

Picture four rubber bands touching and interacting. Now replace the rubber bands with one-dimensional quantum loops, and replace the interaction rules with equations so complex that supercomputers choke on them. That’s the problem of computing string contact interactions in closed string field theory. For decades, the math has fought back.

Closed string field theory (CSFT) describes how closed strings, the tiny loops that form the alphabet of string theory, interact, collide, and exchange energy. It ranks among the most mathematically involved frameworks in physics.

To compute anything useful in CSFT, physicists need to work out the precise geometry of how strings merge and split. For four strings interacting at once, classical numerical methods can just barely manage. For six or more, the problem becomes intractable.

Harold Erbin and Atakan Hilmi Fırat at MIT’s Center for Theoretical Physics have found a way through: train a neural network to do it. Their algorithm solves the geometric problem for four-string interactions and scales, by construction, to any number of strings.

Key Insight: By representing the Strebel quadratic differential, a notoriously difficult geometric object, as a neural network trained with a custom loss function, the researchers bypass the combinatorial explosion that has blocked progress in closed string field theory for decades.

How It Works

The central object is the Strebel quadratic differential: a special mathematical function defined on a sphere with punctures (holes, one per string). Think of it as a map assigning a “flow direction” to every point on the sphere, with punctures acting like sources and sinks. Getting this map right is what allows physicists to compute string amplitudes.

But finding it requires knowing the critical graph, the network of special paths connecting punctures, which you can only determine once you already have the Strebel differential. A classic chicken-and-egg problem.

Classical approaches like Newton’s method handle this for simple cases by guessing the critical graph topology. As the number of punctures grows, the number of possible topologies explodes, and the approach breaks down.

The key advance is a loss function that requires no prior knowledge of the critical graph. The team defined the complex length of a path: an integral measuring how “Strebel-like” a candidate quadratic differential is along any trajectory. When a differential is truly Strebel, all non-contractible paths on the punctured sphere have imaginary part exactly 2π. The loss function penalizes deviations from this condition, pushing training toward the correct solution.

The workflow breaks into three steps:

1. Train the accessory parameter network. The Strebel differential on a 4-punctured sphere depends on a single unknown constant, the accessory parameter, for each arrangement of punctures. A small neural network learns to output this number directly from the puncture coordinates, bypassing iterative root-finding.

2. Extract local coordinates and mapping radii. Once the Strebel differential is known, the researchers expand it around each puncture to find the local coordinates: the precise way to “zoom in” on each string interaction point. The associated mapping radii, which set the physical size of each interaction region, come from numerically evaluating a specific integral.

3. Train the indicator function network. Not every arrangement of four punctures corresponds to a genuine string contact interaction. Some belong to Feynman regions, diagrams built from simpler lower-order interactions that shouldn’t be double-counted. A second neural network learns the indicator function Θ₀,₄: a binary classifier that outputs 1 for configurations in the vertex region (a true contact interaction) and 0 otherwise.

This classifier replaces an explicit geometric description of the vertex region with a trainable function, one that can be retrained for any number of punctures.

To validate the approach, the team computed the 4-tachyon contact term in the tachyon potential. This is a benchmark calculation involving an unstable particle-like state called a tachyon, previously worked out by Nicolas Moeller using classical methods. Their neural network result matched the literature.

Why It Matters

The real payoff extends beyond four strings. Because the algorithm never assumes a particular critical graph topology, it is manifestly independent of the number of punctures. The same framework applies in principle to six, seven, or arbitrarily many punctured spheres. Classical methods fail in exactly this regime, and progress in CSFT has stalled there for years.

Computing interactions on higher-punctured spheres would yield higher-order terms in the tachyon potential, potentially answering a major open question in string theory: does bosonic closed string tachyon condensation produce a stable vacuum? The answer may require sextic-order terms and beyond, which is exactly where this algorithm is headed next.

There’s a broader point here, too. Unsupervised learning with a physics-motivated loss function can solve problems in complex analysis and differential geometry that defeat conventional numerics. Encode the constraint directly into the loss, train without labels, scale by construction. That recipe could work well outside string theory.

Bottom Line: A neural network trained with a graph-agnostic loss function can solve the geometry of string contact interactions, and unlike every previous approach, it scales to as many strings as needed.