Neural Network Generalized Parton Distributions (NNGPD)

This paper presents a deep learning-assisted framework for extracting Generalized Parton Distributions (GPDs) by combining experimental data with ab-initio lattice quantum chromodynamics (LQCD) results to advance the understanding of proton structure.

The Gist

Imagine the proton as a complex, bustling city. For decades, physicists have tried to map this city, but they can't see the streets directly. Instead, they only see the "traffic reports" (experimental data) and the "city planning documents" (theoretical calculations). The goal of this paper is to create a new, super-smart map of the proton's internal structure, known as Generalized Parton Distributions (GPDs).

Here is a simple breakdown of what the authors are doing, using everyday analogies:

The Problem: The "Blind" Map

The proton is made of tiny particles called quarks. To understand how the proton spins and holds together, scientists need to know exactly where these quarks are and how they are moving. This information is the GPD.

However, getting this map is incredibly hard because of two main issues:

1. The Foggy Window (The First Inverse Problem): When scientists look at the proton, they don't see the GPD directly. They see a blurry reflection called a "Compton Form Factor" (CFF). It's like trying to guess the shape of a person standing behind a frosted glass window just by looking at their shadow.
2. The Missing Puzzle Pieces (The Second Inverse Problem): Even if they could see the shadow clearly, turning it back into the original picture is a nightmare. The math involved is like trying to reconstruct a whole cake just by tasting a single crumb. The data is "integrated," meaning the specific details (like the exact position of a quark) are smeared out. Traditional math methods often fail here, leading to many different, conflicting answers (degenerate solutions).

The Solution: The AI Architect

The authors, Zaki Panjsheeri and Simonetta Liuti, have built a new tool called NNGPD (Neural Network Generalized Parton Distributions). Think of this as a highly trained AI architect.

Instead of using rigid, old-school math formulas, they use a Deep Neural Network. This is a computer program modeled after the human brain that learns by example.

Here is how their "AI Architect" works:

• The Training Data: The AI is fed two types of clues:
  1. The "Shadow" (CFFs): Real data from particle accelerators.
  2. The "Blueprints" (Lattice QCD): Super-precise theoretical calculations from supercomputers that act like a ground-truth reference.
• The Rules (Symmetry Constraints): You can't just let the AI guess wildly. The authors programmed it with strict "traffic laws" of physics. For example, the map must look the same if you flip it in certain ways (symmetry). This stops the AI from creating impossible or nonsensical maps.
• The Magic Trick: Traditional methods needed a huge pile of data (like 20+ puzzle pieces) to guess the shape of the proton's interior, and even then, they missed the tiny details at the edges. The authors' AI, however, managed to reconstruct the map accurately using very few pieces of data (only 5 or 6). It's like being able to draw a perfect portrait of a person just by looking at their left ear and a single fingerprint.

The Test: The "Closure Test"

Before using this AI on real, messy experimental data, the authors had to prove it worked. They performed a "closure test."

Imagine they created a fake, perfect proton map (a model called UVA2). They then:

1. Calculated what the "shadows" and "blueprints" would look like for this fake map.
2. Hid the original map.
3. Fed the shadows and blueprints into their AI.
4. Asked the AI to rebuild the map.

The Result: The AI successfully reconstructed the original map almost perfectly. This proves that the framework is capable of solving the puzzle.

The Bottom Line

This paper doesn't claim to have the final map of the proton yet. Instead, it presents a new, powerful framework (the NNGPD) that uses Artificial Intelligence to solve a math problem that has stumped physicists for a long time.

They have shown that by combining experimental data with supercomputer calculations and using a smart, rule-following AI, it is possible to extract a detailed picture of the proton's inner structure with much less data than previously thought possible. The next step, which they note is future work, is to take this framework and apply it to real-world data from actual particle experiments.

Technical Summary

Technical Summary: Neural Network Generalized Parton Distributions (NNGPD)

Problem Statement
Generalized Parton Distributions (GPDs) are fundamental to understanding the 3D spatial structure and angular momentum of the proton. However, extracting GPDs from experimental data presents a "second inverse problem" that is significantly more challenging than extracting standard Parton Distribution Functions (PDFs). While Deeply Virtual Exclusive Scattering (DVES) observables yield Compton Form Factors (CFFs), these CFFs are convolutions of GPDs with Wilson coefficient functions. Crucially, the momentum fraction variable $x$ is integrated out to produce CFFs, leading to a loss of functional dependence on $x$ and a high susceptibility to degenerate solutions. Furthermore, current experimental data quality and the formalism of DVES complicate the initial extraction of CFFs from cross-sections and asymmetries.

Methodology
To address these challenges, the authors propose the NNGPD framework, a deep learning-assisted approach utilizing feed-forward neural networks to construct high-dimensional, bias-controlled representations of GPDs. The methodology integrates three primary constraints:

1. Experimental Data: CFFs extracted from DVES observables.
2. Theoretical Constraints: Mellin moments ($M_n$) and Generalized Form Factors (GFFs) calculated via ab-initio Lattice Quantum Chromodynamics (LQCD).
3. Symmetry Constraints: The framework explicitly enforces QCD symmetries by defining two separate networks for the "plus" ($H^+$) and "minus" ($H^-$) components of the GPDs. The $H^+$ component (antisymmetric in $x$) is constrained by even Mellin moments, while the $H^-$ component (symmetric in $x$) is constrained by odd moments.

The loss function ($L$) minimizes the difference between the "ground truth" and the machine learning-generated CFFs and Mellin moments. Specifically, $L_+$ incorporates CFF constraints, while $L_-$ omits CFF constraints but utilizes odd moments. This architecture is compared to the NNPDF scheme but is distinguished by its incorporation of specific GPD symmetries and the direct use of LQCD moments.

Key Results
The paper reports a closure test performed on the UVA2 phenomenological GPD parameterization, which models valence, sea, and gluon sectors consistent with lattice moments and perturbative QCD evolution.

• Performance: The neural network successfully reconstructed the $H^+_u$ GPD across a wide kinematic range (varying $\xi$, $t$, and $Q^2$) using only $n=5$ to $6$ Mellin moments.
• Efficiency: The study demonstrates that the NNGPD framework, by leveraging inherent symmetries, can accurately recover the $x$-dependence of GPDs with significantly fewer moments than traditional methods (e.g., Bernstein polynomials), which typically require over twenty moments and still struggle to resolve low-$x$ behavior.
• Validation: The results show that the framework can reproduce the input "ground truth" (UVA2) and its uncertainties when trained on data generated from that same model.

Significance and Claims
The authors position NNGPD as a flexible, minimally biased tool for solving the inverse problem of GPD extraction. The paper claims that this framework offers a powerful method to quantitatively answer the proton spin puzzle and provide a comprehensive spatial image of the proton by synergizing experimental CFFs with ab-initio LQCD results.

The work is modest regarding its current scope:

• The results presented are strictly a closure test using a phenomenological model (UVA2); the framework has not yet been trained on actual experimental CFFs or real LQCD data.
• Future work is explicitly identified as the application of this framework to real data and the integration of more interpretable tools, such as symbolic regression, to further refine the solution to the inverse problem.