⚛️ phenomenology

Heavy Quarkonium Spectrum and Decay Constants from a Neural-Network-Based Holographic Model

This paper introduces a neural-network-based holographic model that learns the dilaton field directly from experimental data to successfully reproduce both the heavy quarkonium mass spectrum and the monotonic suppression of leptonic decay constants, overcoming limitations of traditional analytic ansätze.

Original authors: Yu Zhang, Xun Chen, Miguel Angel Martin Contreras

Published 2026-06-19

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Yu Zhang, Xun Chen, Miguel Angel Martin Contreras

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe is built like a giant, invisible 3D hologram. In this hologram, the particles that make up matter (like heavy quarks) are actually shadows cast by a deeper, more complex reality. Physicists call this the "AdS/QCD" model. It's a way to study how these particles stick together without having to solve incredibly difficult math equations directly.

For a long time, scientists trying to map out these particles had to guess the shape of the "holographic background" (the stage on which the particles act). They used simple, pre-made guesses (like assuming the stage was perfectly flat or curved in a simple parabola). But these guesses had a problem: they could predict the weight of the particles well, but they failed to predict how stable they were (how easily they fall apart). It was like having a map that got the cities right but got the roads between them completely wrong.

The New Approach: Teaching a Computer to "Feel" the Shape

In this paper, the authors decided to stop guessing. Instead, they used a Neural Network (a type of artificial intelligence) to figure out the shape of the stage directly from real-world data.

Think of it like this:

The Old Way: A sculptor tries to carve a statue of a horse by following a textbook drawing. The result looks okay, but the legs are a bit stiff, and the muscles don't move naturally.
The New Way: The sculptor puts the horse in front of a smart camera (the Neural Network). The camera watches the horse run, jump, and rest. The AI learns the exact curve of every muscle and bone just by watching the real animal. It doesn't follow a textbook; it follows the data.

How They Did It

The Input: They fed the AI data from the Particle Data Group (PDG), which is like the "encyclopedia" of particle physics. They gave it the known weights and stability levels of heavy particles called Charmonium (made of charm quarks) and Bottomonium (made of bottom quarks).
The Learning: The AI tried to draw a smooth curve (called the "dilaton field") that would explain why these particles have the weights and stability they do. It used a special trick called "automatic differentiation" to check its own work instantly, adjusting the curve until it fit the data perfectly.
The Result: The AI discovered that the shape of the universe's "stage" isn't a simple curve. It's a complex, wavy shape that changes depending on how far you are from the center.
- Near the center (UV): The shape is slightly different from what old theories predicted. This small change is crucial because it explains why heavier, excited particles become less stable (their "decay constants" drop).
- Far away (IR): The shape grows rapidly, acting like a tight rubber band that keeps the particles stuck together (confinement).

Why This Matters

The old models were like a pair of glasses that were blurry on one side. They could see the mass of the particles clearly, but the stability was fuzzy. The new AI-generated model puts on a fresh pair of glasses that are sharp on both sides.

Accuracy: The new model predicted the masses of these particles with an error of only about 1.26% for Charmonium and 3.32% for Bottomonium. That's incredibly precise.
Solving a Mystery: For years, physicists struggled to explain why heavier versions of these particles get less stable as they get "excited" (like a guitar string vibrating faster). The AI found a specific shape for the background field that naturally causes this drop in stability, solving a puzzle that had stumped researchers for a long time.

A Lesson on "Guessing" with AI

The authors also tested different "brain settings" for their AI (called activation functions). They found that if they used a setting that allowed the AI to grow wildly unbounded (like a ReLU function), the AI would make wild, unrealistic guesses for particles it hadn't seen before. However, if they used a "bounded" setting (like Tanh), the AI was forced to be more conservative and realistic, acting like a built-in safety guard. This taught them that in science, the type of math you choose for your AI is just as important as the data you feed it.

In Summary

This paper shows that by letting a computer learn the shape of the universe's "holographic stage" directly from experimental data, we can finally get a perfect picture of how heavy particles behave. It's a move from "guessing the rules" to "learning the rules from the players," resulting in a much more accurate and unified theory of how these particles stick together and fall apart.

Technical Summary: Heavy Quarkonium Spectrum and Decay Constants from a Neural-Network-Based Holographic Model

Problem Statement
Traditional bottom-up AdS/QCD models, particularly the soft-wall model, rely on ad hoc analytic ansätze for the dilaton field $\Phi(z)$ , typically assuming a quadratic form ( $\Phi(z) \propto z^2$ ). While this successfully reproduces linear Regge trajectories for light mesons, it fails to simultaneously describe the heavy quarkonium (charmonium and bottomonium) spectrum and their leptonic decay constants. Specifically, the quadratic ansatz predicts degenerate decay constants or an incorrect trend, failing to capture the experimentally observed monotonic suppression of decay constants with increasing radial excitation. Furthermore, the linear Regge trajectory hypothesis ( $M_n^2 \propto n$ ) breaks down for heavy quark systems, which require non-linear parametrizations ( $M_n^2 \propto (n+b)^\nu$ ) consistent with Bethe-Salpeter theory and non-relativistic QCD dynamics. Existing attempts to resolve these issues via finite UV cutoffs or modified potentials have yielded only qualitative agreement or failed to fit both mass spectra and decay constants simultaneously with high precision.

Methodology
The authors propose a data-driven inverse construction of the dilaton field $\Phi(z)$ using a Multilayer Perceptron (MLP) within the bottom-up AdS/QCD framework.

Neural Network Architecture: The model employs an MLP with five hidden layers (128 neurons each) to learn the derivative of the dilaton field, $\Phi'(z)$ , as a smooth function of the holographic coordinate $z$ . The network output is structurally constrained to satisfy the ultraviolet (UV) boundary condition $\Phi(0)=0$ , ensuring consistency with the conformal dimension of the vector current operator.
Automatic Differentiation: Instead of fitting the potential directly, the network outputs $\Phi'(z)$ . The second derivative $\Phi''(z)$ is computed via automatic differentiation. These derivatives are substituted into the holographic potential equation:
$U(z) = \frac{3}{4z^2} + \frac{\Phi'(z)}{2z} + \frac{\Phi'(z)^2}{4} - \frac{\Phi''(z)}{2}$
Spectral Calculation: The resulting potential $U(z)$ is used in a Schrödinger-like equation. This equation is discretized on a finite grid and diagonalized to extract eigenvalues (masses $M_n$ ) and eigenfunctions. Decay constants ( $f_n$ ) are calculated from the near-boundary behavior of the bulk-to-boundary modes.
Training Strategy: The network is trained on experimental data from the Particle Data Group (PDG) for charmonium and bottomonium states. The objective is to minimize the root-mean-square (RMS) deviation between predicted and experimental masses and decay constants. The authors utilize the Adam optimizer with a cosine-annealing learning rate schedule.
Activation Function Analysis: A critical component of the methodology involves testing different activation functions (Tanh, ReLU, SiLU). The study demonstrates that bounded, saturating functions (like Tanh) are essential for preventing unphysical extrapolation in data-scarce regimes, acting as an implicit regularizer.

Key Contributions

First MLP Application to Dilaton Reconstruction: This work represents the first application of MLPs to directly reconstruct the dilaton field $\Phi(z)$ in a bottom-up AdS/QCD model, moving away from fixed analytic ansätze.
Simultaneous Fit of Masses and Decay Constants: The model successfully resolves the long-standing difficulty of simultaneously reproducing the heavy quarkonium mass spectrum and the monotonic decrease of leptonic decay constants with radial excitation.
Non-Quadratic Dilaton Profile: The data-driven reconstruction yields a non-quadratic dilaton profile. In the UV region, it exhibits non-quadratic corrections that lift the degeneracy of decay constants; in the IR region, rapid growth reflects confinement and produces non-linear Regge trajectories.
Inductive Bias in Activation Functions: The paper identifies that the choice of activation function is a decisive component of the model's inductive bias. It shows that Tanh activation functions provide superior stability and accuracy for sparse hadron spectroscopy data compared to unbounded functions like ReLU, which suffer from uncontrolled extrapolation.

Results

Accuracy: The combined fit achieves RMS deviations of 1.26% for charmonium and 3.32% for bottomonium.
Mass Spectrum: The model predicts masses for ground and excited states with high precision. For example, the predicted mass for $J/\psi$ deviates by only 0.01% from the experimental value, and $\Upsilon(1S)$ by 0.0002%.
Decay Constants: The model correctly reproduces the experimental trend where decay constants decrease as the radial excitation number $n$ increases, a feature previous soft-wall models failed to capture quantitatively.
Generalization: When trained on a subset of states (e.g., the first three charmonium states) and tested on higher excitations, the model maintains reasonable predictive power, though deviations increase for the highest excited states (e.g., $\Upsilon(11020)$ shows a 1.3% mass deviation and 11.3% decay constant deviation).
Potential Structure: The reconstructed effective potentials $U(z)$ show near-harmonic oscillator behavior in the UV and necessary non-linear corrections in the IR, consistent with the physical requirements for heavy quark systems.

Significance
The paper establishes neural-network reconstruction as a flexible and effective tool for holographic modeling, providing a unified framework that captures the essential dynamics of heavy quarkonia without relying on ad hoc parametrizations. The work demonstrates that the suppression of the wave function at the origin (governing decay constants) and the confinement mechanism (governing the spectrum) can be unified within a single holographic potential derived directly from data. The authors claim this approach offers a basis for future extensions, such as incorporating finite-temperature backgrounds, lattice QCD constraints, or additional channels, while emphasizing that the current results are primarily focused on the zero-temperature spectrum. The study also highlights the importance of activation function design as a form of implicit regularization in physics-informed neural networks dealing with sparse data.

Technical Summary: Heavy Quarkonium Spectrum and Decay Constants from a Neural-Network-Based Holographic Model

More like this