Physics-Informed Global Extraction of the Universal Small-$x$ Dipole Amplitude

This paper presents a physics-informed neural network approach to extract a universal small-$x$ dipole scattering amplitude directly from global DIS and photoproduction data without rigid parametric assumptions, successfully resolving the long-standing tension between total and charm cross-section channels while providing a smooth, non-negative input for Color Glass Condensate phenomenology.

The Gist

Imagine you are trying to understand the inside of a giant, chaotic storm cloud. In the world of particle physics, this "storm" is a proton (the building block of atoms) being hit by a high-speed particle. Inside this storm, there are tiny particles called gluons (the "glue" holding the proton together) that are moving so fast and are so numerous that they start to overlap and merge, creating a dense, saturated state of matter. Physicists call this the Color Glass Condensate.

The paper you shared is like a new, super-smart weather map that finally gets the shape of this storm cloud right, solving a puzzle that has confused scientists for years.

Here is the breakdown using simple analogies:

1. The Problem: The "Rigid Mold" vs. The "Real Shape"

For a long time, scientists tried to describe this proton storm using a rigid mold (a fixed mathematical formula). They would say, "We think the storm looks like a perfect sphere," or "It looks like a specific type of egg."

• The Issue: When they tried to fit this rigid mold to real data, it worked okay for some parts of the storm (like the total energy) but failed miserably for other parts (like the heavy "charm" particles). It was like trying to fit a square peg into a round hole.
• The Conflict: The data from the "total energy" measurements and the "charm particle" measurements were fighting each other. One said the storm was one shape; the other said it was another.

2. The Solution: The "Physics-Informed AI"

Instead of forcing the data into a rigid mold, the authors used a Physics-Informed Neural Network (PINN).

Think of a standard AI as a student who just memorizes flashcards of past weather reports. It might guess the weather, but it doesn't understand why it rains.
This new AI is different. It's like a student who memorizes the flashcards AND knows the laws of physics (like gravity and wind pressure).

• The "Teacher" (The Laws): The AI is taught the BK Equation, which is the rulebook for how gluons behave and evolve at high speeds. The AI must follow these rules.
• The "Homework" (The Data): The AI is also shown real photos of the storm from giant particle colliders (like the HERA lab).
• The Result: The AI doesn't just guess a shape; it learns the shape of the storm that fits both the photos and the laws of physics perfectly. It's flexible, like clay, rather than rigid like a plastic mold.

3. The Three Big Wins

This new method solved three major headaches:

• Win #1: The "Charm" Problem Solved.
  Previously, the AI couldn't explain why heavy charm particles were behaving the way they were. By letting the shape of the storm be flexible, the new model finally fits both the total energy data and the charm particle data at the same time. It's like finally finding a single weather map that explains both the rain and the wind.

• Win #2: The "Ghost" Problem Solved.
  In the old models, when scientists tried to look at the storm from a different angle (using math called a "Fourier transform"), they sometimes saw "ghosts"—negative numbers where there should be nothing. In physics, you can't have "negative probability" or "negative glue."
  The new AI was given a strict rule: "No ghosts allowed." It learned to shape the storm so that when you look at it from any angle, everything stays positive and real.

• Win #3: A Universal Map.
  The result is a single, smooth, universal map of the proton's interior. This map works for:

  • Smashing protons together (like at the LHC).
  • Shooting electrons at protons (like the future Electron-Ion Collider).
  • Creating heavy particles like the J/psi (a specific type of particle).
    Before, scientists needed different maps for different experiments. Now, they have one master map that works for everything.

The Bottom Line

This paper is a breakthrough because it stopped trying to force nature into a box. Instead, it built a smart, flexible AI that respects the laws of physics while learning from real data.

The Analogy:
Imagine trying to draw a portrait of a person who is constantly moving.

• Old Method: You tried to draw them using a stencil (a fixed shape). It looked okay when they stood still, but when they moved, the drawing looked wrong.
• New Method: You used a smart camera that knows how human muscles work (the physics) and takes thousands of photos (the data). It creates a 3D model that moves naturally, looks real in every photo, and never has "glitchy" parts.

This new "3D model" of the proton will help physicists understand the universe better, especially as they build bigger, more powerful particle colliders in the future.

Technical Summary

Here is a detailed technical summary of the paper "Physics-Informed Global Extraction of the Universal Small-$x$ Dipole Amplitude" by Dai et al.

1. Problem Statement

The paper addresses critical challenges in the global analysis of small-$x$ (Bjorken-$x$) Quantum Chromodynamics (QCD) within the Color Glass Condensate (CGC) effective field theory. Specifically, it targets three long-standing issues in determining the universal dipole scattering amplitude, $N(r, x_B)$:

• Tension between Observables: Conventional fits based on rigid parametric ansätze (e.g., MV-type models) struggle to simultaneously describe the reduced total deep inelastic scattering (DIS) cross-section ($\sigma_r$) and the reduced charm cross-section ($\sigma_{c\bar{c}}^r$). Charm production probes smaller dipole sizes, imposing stringent constraints on the short-distance behavior of $N(r, x_B)$ that standard parametrizations often fail to satisfy without compromising the fit to inclusive data.
• Fourier-Positivity Violation: Many widely used initial conditions for the dipole amplitude, when Fourier-transformed to momentum space to yield the dipole $S$-matrix $S(k_T, x_B)$, produce negative values in certain transverse-momentum ($k_T$) regions. This violates physical requirements (Fourier positivity) and complicates the interpretation of the amplitude as a gluon density.
• Parametrization Bias: Traditional approaches impose a specific analytic functional form for the non-perturbative initial condition at a starting scale $x_0$. This introduces model bias and limits the flexibility needed to capture complex dynamics across different kinematic regimes.

2. Methodology: Physics-Informed Neural Networks (PINN)

The authors propose a novel framework using Physics-Informed Neural Networks (PINN) to extract the dipole amplitude $N(r, x_B)$ as a smooth, differentiable surrogate function without imposing a priori parametric forms.

• Network Architecture: A Residual Neural Network (ResNet) with seven residual blocks takes the dipole size $r$ and rapidity $Y = \ln(x_0/x_B)$ as inputs. The output is constrained to the physical range $(0, 1)$ via a sigmoid activation function.

• Loss Function: The training objective is a weighted sum of three components:
  $$\mathcal{L}(\theta, p) = w_1 \mathcal{L}_{\text{ciBK}} + w_2 \mathcal{L}_{\text{data}} + w_3 \mathcal{L}_{\text{phy}}$$

  1. Physics Loss ($\mathcal{L}_{\text{ciBK}}$): Enforces the collinearly improved Balitsky–Kovchegov (ciBK) evolution equation. The network is penalized for the residual of the integro-differential equation, ensuring the solution satisfies QCD evolution dynamics directly.
  2. Data Loss ($\mathcal{L}_{\text{data}}$): A negative log-likelihood term measuring agreement with experimental data. This includes:
     • HERA reduced total cross-section ($\sigma_r$) at $\sqrt{s} = 224.9, 251.5, 300.3, 318$ GeV.
     • HERA reduced charm cross-section ($\sigma_{c\bar{c}}^r$) at $\sqrt{s} = 318$ GeV.
     • Exclusive $J/\psi$ photoproduction data from ALICE, H1, ZEUS, and LHCb.
     • Uncertainties are quantified using Deep Ensembles (training multiple independent networks on pseudodata replicas).
  3. Physical Constraints Loss ($\mathcal{L}_{\text{phy}}$): Enforces fundamental QCD principles:
     • Black-disk limit: $N(r_{\text{max}}, Y) \to 1$ for large dipoles.
     • Color transparency: Constrains the local logarithmic slope $\partial \ln N / \partial \ln r^2$ to be between 0 and 1 for small dipoles.
     • Fourier Positivity: Explicitly penalizes negative values of the momentum-space $S$-matrix $S(k_T, Y) = \int d^2r e^{i k_T \cdot r} [1 - N(r, Y)]$.

• Optimization: The network parameters $\theta$ and physical parameters (effective proton size $R_p$, normalization factor $K_{VM}$, infrared-frozen coupling $\alpha_{fr}$) are optimized simultaneously using the AdamW algorithm with adaptive weight balancing.

3. Key Contributions

• Non-Parametric Extraction: The first global analysis to infer the initial dipole profile non-parametrically from data, constrained only by the ciBK evolution equation and physical laws, rather than a fixed analytic ansatz.
• Resolution of the $\sigma_r$ vs. $\sigma_{c\bar{c}}^r$ Tension: By allowing the functional form to adapt freely, the PINN successfully describes both inclusive and heavy-flavor DIS data simultaneously, a task where previous rigid models failed.
• Enforcement of Momentum-Space Positivity: The study demonstrates that a smooth, non-negative momentum-space dipole can be achieved globally by incorporating a Fourier-positivity penalty, resolving a persistent issue in CGC phenomenology.
• Unified Framework: Provides a single, universal amplitude $N(r, x_B)$ that consistently describes inclusive DIS, heavy-flavor production, and exclusive vector-meson photoproduction ($J/\psi$).

4. Results

• Global Fit Quality:
  • The PINN solution achieves a $\chi^2/\text{d.o.f.} = 1.150$ for the reduced total cross-section ($\sigma_r$) and $1.546$for the reduced charm cross-section ($\sigma_{c\bar{c}}^r$).
  • It describes exclusive $J/\psi$ photoproduction with an excellent $\chi^2/\text{d.o.f.} = 0.474$, validating the universality of the extracted amplitude across inclusive and exclusive channels.
• Amplitude Characteristics:
  • The extracted $N(r, x_B)$ is smooth and differentiable.
  • Comparison with a standard MV-type parametrization fitted to the PINN output at $x_B=0.01$ reveals systematic deviations, particularly at small $r$, indicating that rigid MV forms cannot capture the true structure required by the data.
  • The effective anomalous dimension is determined locally from the network gradients rather than being a fixed constant.
• Momentum Space:
  • The Fourier-transformed $S$-matrix, $S(k_T, x_B)$, remains strictly non-negative over the full examined transverse-momentum range ($k_T \in [0, 100]$ GeV) for $x_B \le 0.01$. This contrasts with previous Bayesian fits that exhibited oscillations or negative regions.
• Kinematic Constraints: The analysis is restricted to $x_B < 0.004$ to avoid the transition region where DGLAP dynamics overlap with saturation, ensuring the validity of the pure small-$x$ evolution framework used.

5. Significance

This work represents a paradigm shift from "data-driven" fitting to "physics-driven" inference in high-energy nuclear physics.

• Robust Input for CGC: The extracted amplitude provides a well-behaved, positive-definite input for future CGC phenomenology, essential for calculating particle production in $e$–$p/A$ and $p$–$p/A$ collisions.
• Preparation for EIC: The methodology is ideally suited for the upcoming Electron-Ion Collider (EIC), which will provide high-precision data over a broad kinematic range. The PINN framework can naturally incorporate these new constraints without re-parameterizing the model.
• Generalizability: The approach offers a robust template for scientific inference in other domains where dynamical equations (PDEs) and first-principles constraints coexist with experimental measurements.

The authors have made the tabulated PINN solution and its Fourier transform publicly available, facilitating immediate application in downstream theoretical calculations.