Extraction of the color dipole amplitude with physics-informed neural networks

This paper introduces a Physics-Informed Neural Network framework using a "Teacher–Student" strategy to extract a model-independent color dipole amplitude from HERA data, which successfully predicts exclusive $J/\psi$ photoproduction cross-sections without parameter retuning, thereby providing strong evidence for the process-independence of the gluon saturation scale in high-energy QCD.

The Gist

Imagine the proton (a tiny particle inside an atom) not as a solid marble, but as a bustling, chaotic city filled with invisible messengers called gluons. When you zoom in really close and look at these gluons moving at incredibly high speeds, they multiply so fast that they start to crowd each other, forming a dense, saturated "traffic jam." Physicists call this state the Color Glass Condensate.

The paper you provided is about figuring out exactly how dense this traffic jam is and how it behaves, using a new kind of "smart detective" tool.

Here is the breakdown of their work in simple terms:

The Problem: The "Rigid Map" vs. The "Real City"

For a long time, scientists tried to map this gluon traffic jam using a "rigid map." They would guess a shape for the traffic jam (a mathematical formula) and then tweak the numbers until it fit the data from one type of experiment (called inclusive experiments, where they smash particles and look at the general debris).

However, when they tried to use that same map to predict a different type of experiment (called exclusive experiments, where they look for a specific, rare particle called a J/ψ meson popping out), the map failed. To make it work, they had to manually stretch or shrink the map (geometric adjustments) just to make the numbers match. It was like trying to use a flat map of a city to navigate a mountain; it didn't work without forcing the terrain to fit the paper.

The Solution: The "Teacher-Student" AI

The authors, Wei Kou and Xurong Chen, introduced a new method using Physics-Informed Neural Networks (PINNs). Think of this as a two-person team solving a mystery:

1. The Teacher (The Physics Rules): This is the "Teacher." It knows the fundamental laws of how gluons behave (specifically an equation called the Balitsky-Kovchegov or BK equation). It doesn't care about the messy data yet; it just knows the rules of the game. It says, "The traffic jam must evolve in this specific way according to the laws of physics."
2. The Student (The Data Learner): This is the "Student." It looks at the actual experimental data from the HERA accelerator (real-world observations of the proton). Its job is to learn what the traffic jam actually looks like based on what the sensors saw.

How they work together:
The "Teacher" constantly checks the "Student's" work. If the Student tries to draw a traffic jam that breaks the laws of physics, the Teacher corrects it. If the Student ignores the real data, the Teacher pushes it back toward the observations.

The result is a universal map of the gluon traffic jam. Crucially, they didn't have to guess the starting shape of the jam or force it to fit. The AI learned the shape directly from the data while obeying the laws of physics.

The Big Surprise: One Map Fits All

Usually, a map that fits one type of experiment fails at another. But here is the magic of their discovery:

They trained their AI using only the "inclusive" data (the general debris). They then took that exact same map and used it to predict the "exclusive" data (the rare J/ψ particle).

They did not change a single number. They didn't tweak the map or stretch it. They just handed the map to the exclusive experiment, and it worked perfectly.

Why This Matters

This is a huge deal because it proves that the "gluon saturation scale" (the point where the traffic jam gets so dense it stops growing) is universal. It behaves the same way regardless of how you look at it.

• The Analogy: Imagine you learn to drive a car by practicing in a parking lot (inclusive data). Usually, you might think, "I'm great at parking, but I'd crash on a highway." But this paper shows that if you truly understand the laws of driving (the physics), you can drive on the highway (exclusive data) perfectly without needing to re-learn how to steer.

The Bottom Line

The authors successfully used a "Teacher-Student" AI to extract a pure, unbiased picture of how gluons behave inside a proton. They showed that this picture is so accurate and fundamental that it can predict complex, rare particle events without any extra adjustments. This suggests that the underlying rules of the strong force (which holds atoms together) are consistent and universal, and that this new AI approach is a powerful way to uncover those hidden laws.

Technical Summary

Technical Summary: Extraction of the Color Dipole Amplitude with Physics-Informed Neural Networks

Problem Statement
The extraction of the universal color-dipole scattering amplitude, a cornerstone of high-energy Quantum Chromodynamics (QCD) and the Color Glass Condensate (CGC) framework, presents a significant inverse problem. While the Balitsky-Kovchegov (BK) equation rigorously describes the non-linear rapidity evolution of this amplitude, standard phenomenological approaches struggle to reconcile inclusive Deep Inelastic Scattering (DIS) data with exclusive (diffractive) measurements. Traditional methods typically rely on rigid parametric ansatzes and often necessitate ad-hoc geometric adjustments to fit diverse datasets. Furthermore, extracting the amplitude directly from data without assuming specific initial states remains a challenge, particularly in testing the robustness of the saturation picture and the process-independence of the gluon saturation scale.

Methodology
The authors introduce a Physics-Informed Neural Network (PINN) framework to resolve these tensions by embedding the momentum-space BK evolution dynamics directly into the learning process. The methodology employs a "Teacher–Student" training strategy:

1. Physics Constraint (Teacher): The network is initially constrained by the BK equation in the diffusion approximation. The partial differential equation (PDE) governing the evolution of the dipole amplitude $N(k, Y)$ in momentum space acts as a "teacher," defining the physical solution manifold before the network encounters experimental data. The PDE constraint is formulated as:
   $$\frac{\partial N}{\partial Y} = \bar{\alpha}_s \left( A_0 N - A_1 \frac{\partial N}{\partial L} + A_2 \frac{\partial^2 N}{\partial L^2} - N^2 \right)$$
   where $L = \ln(k^2/\Lambda_{QCD}^2)$ and $Y = \ln(1/x)$.

2. Data Refinement (Student): A fully connected neural network (FCNN) with four hidden layers (64 neurons each) and a Sigmoid output activation (to enforce the unitarity bound $0 \le N \le 1$) acts as the "student." It is refined against inclusive HERA $F_2$ data. The total loss function combines the PDE residual ($L_{PDE}$), the data mismatch ($L_{Data}$), and boundary conditions ($L_{Bound}$).

3. Direct Momentum-Space Calculation: To avoid numerical instabilities associated with Fourier-Bessel transforms during training, the inclusive structure function $F_2(x, Q^2)$ is computed via direct convolution of the momentum-space photon wavefunctions and the dipole amplitude $N(k, Y)$.

4. Uncertainty Quantification: The authors utilize Bootstrap Aggregating (Bagging) with an ensemble of 20 independent PINNs trained on resampled datasets to quantify epistemic uncertainty, reporting results as mean $\pm 2\sigma$.

Key Contributions and Results

• Model-Independent Extraction: The framework successfully extracts a model-independent dipole amplitude without assuming specific initial states or rigid parametric forms. The network learns the effective coefficients of the diffusion approximation ($A_0, A_1, A_2$) directly from the data, yielding a growth rate $A_0 \approx 0.661$ and an effective saturation exponent $\lambda_s \approx 0.239$, consistent with phenomenological extractions.
• High-Fidelity Inclusive Fit: The PINN achieves a global fit to HERA $F_2$ data with a $\chi^2/\text{dof} \approx 0.972$. The model accurately captures the kinematic dependence across the entire $(x, Q^2)$ plane, with uncertainty bands narrowing in the high-$k$ (perturbative) regime due to the regularization effect of the PDE loss.
• Parameter-Free Prediction of Exclusive Processes: A critical result is the prediction of exclusive $J/\psi$ photoproduction cross-sections. Using the dipole amplitude and geometric parameters (proton radius $R_p \approx 5.46 \text{ GeV}^{-1}$) derived solely from the inclusive $F_2$ fit, the model predicts the absolute normalization and kinematic dependence of exclusive $J/\psi$ production without any further retuning or geometric rescaling.
• Emergent Dynamics: The trained solution spontaneously exhibits geometric scaling and the propagation of a saturation wavefront in momentum space, emerging naturally from the interplay of linear growth and non-linear damping terms learned by the network.

Significance and Claims
The paper claims that this work establishes a transformative paradigm for uncovering non-perturbative QCD structures. By demonstrating that a dipole amplitude constrained by BK dynamics and fitted only to inclusive data can successfully describe exclusive observables without additional tuning, the authors provide compelling evidence for the process-independence of the gluon saturation scale.

The authors position PINNs not merely as fitting tools but as a rigorous method for solving ill-posed inverse problems in QCD thermodynamics and spectroscopy. They note that while the current results are limited by the fixed-coupling diffusion approximation (valid primarily for small $x$ and moderate $Q^2$), the framework successfully bridges perturbative evolution and non-perturbative hadronic structure, resolving the long-standing tension between inclusive and diffractive measurements without ad-hoc adjustments. The work suggests that future iterations incorporating running coupling and next-to-leading order (NLO) corrections could extend this precision to broader kinematic windows, particularly for upcoming Electron-Ion Collider (EIC) data.