Reconstruction of Gravitational Form Factors using Generative Machine Learning

This paper introduces a generative framework based on denoising diffusion to perform model-independent, non-parametric reconstruction of proton gravitational form factors from sparse and noisy data, enabling robust extraction of chiral low-energy constants and the nucleon D-term across the full kinematic range.

The Gist

Imagine you are trying to draw a perfect picture of a mountain range, but you only have a few blurry, shaky photos taken from a distance, and some of them are covered in static. You know the general shape of mountains (they have peaks, valleys, and slopes), but you don't know the exact details of this specific range.

This is exactly the problem physicists face when studying the proton (the particle inside atoms that gives matter its mass). They want to map out the "gravitational form factors" of the proton. Think of these as a 3D map showing how mass, spin, and pressure are distributed inside a proton.

The problem? The data they have is sparse (very few points) and noisy (full of errors). Traditionally, to fill in the gaps, scientists had to guess a specific mathematical shape (like a curve or a wave) to connect the dots. But if they guessed the wrong shape, the whole map could be wrong.

This paper introduces a new, smarter way to solve this puzzle using Generative AI, specifically a type called a Denoising Diffusion Model.

Here is how it works, broken down into simple analogies:

1. The "Art School" Training (The Prior)

Before the AI can help, it needs to learn what a "real" proton looks like. The researchers didn't just feed it random lines. They created a massive "art school" for the AI.

• They generated 600,000 synthetic curves based on ten different, well-respected theories of how protons work (like Chiral Perturbation Theory, Lattice QCD, and others).
• The Analogy: Imagine teaching an artist to draw mountains not by showing them one specific mountain, but by showing them thousands of photos of every type of mountain on Earth—volcanoes, snow-capped peaks, rolling hills, and jagged cliffs. The AI learns the rules of mountain shapes without being told to draw a specific one.

2. The "Denoising" Process (The Magic)

The AI uses a technique called Denoising Diffusion.

• The Analogy: Imagine you have a clear photo of a mountain, but someone slowly adds fog and static to it until it's just white noise. The AI is trained to reverse this process. It learns to look at a blurry, noisy mess and "peel back" the fog step-by-step to reveal the clear image underneath.
• The Twist: In this paper, the AI doesn't just start with noise. It starts with noise and a few specific data points (the blurry photos the physicists actually have). It is forced to generate a mountain that fits those few known points while still looking like a "real" mountain based on its training.

3. The Result: A Map Without Guessing

When they applied this to the proton's gravitational data:

• The Magic: The AI successfully reconstructed the entire map of the proton's internal forces, even when they only gave it one or two data points to work with.
• Why it's special: Because the AI learned the "physics" of the proton from its training, it didn't need to guess a mathematical formula. It just "knew" that the curve had to behave a certain way. It filled in the gaps with a smooth, physically realistic curve that fits the sparse data perfectly.

4. The Big Discovery: The "D-Term"

The most exciting part of the paper is what they found about the D-term.

• The Analogy: Think of the proton as a tiny, spinning ball of jelly. The "D-term" is a measure of how much internal pressure is pushing the jelly apart versus holding it together. It's a fundamental property, just as important as the proton's mass or charge.
• The Problem: Scientists have been struggling to measure this because the data is so noisy near the center.
• The Solution: Using their new AI method, they calculated the D-term with high precision. They found a value of -4.3.
• The Check: This number matches perfectly with other independent methods (like a "dispersive" method that uses scattering data), proving that their AI approach is reliable.

5. Why This Matters

• Efficiency: In the past, to get a good map, you needed a mountain of data. Now, with this AI, you can get a high-quality map with just a few data points. This saves massive amounts of computing power and time.
• Reliability: It removes the bias of "guessing the shape." The AI lets the data and the laws of physics speak for themselves.
• Future: This method can be used for other particles (like the Delta baryon or rho-mesons) and even for nuclear systems. It's like giving physicists a new, super-powerful microscope that can see the invisible structure of matter with very little light.

In a nutshell: The authors built a smart AI that learned the "rules of the game" for how protons behave. When given a few blurry clues, it used those rules to reconstruct the entire, clear picture of the proton's internal gravity, solving a decades-old puzzle about its internal pressure without needing to guess the answer.

Technical Summary

Here is a detailed technical summary of the paper "Reconstruction of gravitational form factors using generative machine learning" by Herzallah Alharazin and Julia Yu. Panteleeva.

1. Problem Statement

The paper addresses the challenge of reconstructing hadronic Gravitational Form Factors (GFFs)—specifically $A(t)$, $J(t)$, and $D(t)$ for the proton—from sparse and noisy data.

• Context: GFFs describe the spatial distribution of energy, momentum, spin, pressure, and shear forces inside hadrons. They are crucial for understanding nucleon structure but are difficult to measure directly.
• Limitations of Current Methods:
  • Lattice QCD: While powerful, current calculations suffer from systematic uncertainties (single gauge ensemble, unphysical pion masses) and intrinsically noisy signals, particularly near $t=0$. Extrapolating to $t=0$ (the D-term) is highly model-dependent.
  • Chiral Perturbation Theory (ChPT): Valid at low energies but relies on unknown Low-Energy Constants (LECs) ($c_8, c_9$) that must be determined from data.
  • Parametric Fits: Traditional approaches impose specific functional forms (e.g., dipole, $z$-expansion) on the $t$-dependence, introducing model bias and potentially missing non-analytic behaviors.
• Goal: Develop a model-independent framework that can robustly reconstruct the full kinematic range of GFFs using very few data points, without assuming a specific functional ansatz for the $t$-dependence.

2. Methodology

The authors propose a physics-guided Denoising Diffusion Probabilistic Model (DDPM) framework.

A. Generative Prior Construction

Instead of learning from raw experimental data, the model is trained on a large ensemble of synthetic curves ($6 \times 10^5$ samples) representing physically plausible GFF shapes.

• Data Source: Ten distinct functional classes were used to ensure diversity:
  1. Multipole: Standard lattice parametrization.
  2. $z$-expansion: General analytic parametrization.
  3. Meson Dominance: $t$-channel meson exchanges.
  4. Modified Exponential: Constituent quark/AdS/QCD shapes.
  5. Padé Approximants: Rational functions.
  6. Dispersive: Based on analyticity and spectral functions.
  7. Log-modified Multipole: Incorporating QCD running coupling corrections.
  8. Bag Model: Bessel-type forms from MIT bag model.
  9. Convex Combinations (Pairwise): Mixing the above classes.
  10. Dirichlet Mixtures: Randomized mixtures of multiple classes.
• Training: The model learns the manifold of these shapes, ensuring the prior covers the full space of analytic functions without privileging any single parametrization.

B. The Diffusion Architecture

• Model: A 1D ResNet-Attention hybrid network using $v$-prediction (predicting the velocity vector rather than noise or clean data) to ensure uniform variance across noise levels.
• Conditioning Mechanism: The reconstruction is a conditional generation problem (analogous to image inpainting).
  • Concatenation: Known data points are mapped to the grid and concatenated with a mask and the noisy estimate.
  • Replacement: At each reverse diffusion step, known grid points are overwritten with a noisy version of the conditioning values to enforce local consistency.
  • Uncertainty Propagation: Experimental errors in the conditioning points are injected as noise ($\xi_n$) into the conditioning vector for each generated sample, ensuring the final output reflects both model uncertainty and data error.

C. Physics Constraints

The framework incorporates physical constraints directly:

• Poincaré Algebra: Exact constraints $A(0)=1$ and $J(0)=1/2$ are enforced as conditioning points.
• Sign Constraints: For the $D(t)$ form factor, the physical requirement of mechanical stability ($D(0) < 0$) is enforced by filtering the generated ensemble.

3. Key Contributions

1. Model-Independent Reconstruction: The first application of diffusion models to hadronic form factors, achieving non-parametric reconstruction that does not rely on predefined functional forms (like dipoles or $z$-expansions).
2. Robustness to Data Scarcity: Demonstrated that the framework yields stable, accurate reconstructions even when conditioned on only one or two data points, provided the physical prior is strong.
3. Extraction of Low-Energy Constants: Enabled the direct extraction of the unknown ChPT constants $c_8$ and $c_9$ by matching the diffusion model's output to ChPT expressions in the low-$|t|$ region.
4. Prediction of the Nucleon D-term: Provided a robust, model-independent estimate of the nucleon D-term, $D(0)$, at the physical pion mass.

4. Key Results

A. Reconstruction of $A(t)$ and $J(t)$ (Unphysical Pion Mass)

• Using lattice QCD data from Ref. [17], the model reconstructed $A(t)$ and $J(t)$ across $0 \le -t \le 2 \text{ GeV}^2$.
• Data Ablation Study: Even with only one intermediate lattice point plus the $t=0$ constraint, the median reconstruction remained consistent with standard parametric fits and the full lattice dataset. The credible interval widened only in regions where data was removed, providing a data-driven measure of extrapolation uncertainty.

B. Reconstruction of $D(t)$ (Unphysical Pion Mass)

• $D(t)$ is the most challenging due to the lack of a sum-rule constraint at $t=0$ and large lattice errors.
• Enforcing $D(0) < 0$ (mechanical stability) was sufficient to constrain the entire curve.
• The model successfully reconstructed the quark and gluon contributions to $D(t)$ using only two lattice points, matching tripole parameterizations without assuming that form.

C. Extraction of Low-Energy Constants (LECs)

By matching the diffusion model's posterior to ChPT expressions at the unphysical pion mass:

• $c_9 = -0.61 \pm 0.19 \text{ GeV}^{-1}$ (extracted from $A(t)$ and $J(t)$).
• $c_8 = -4.6 \pm 0.8 \text{ GeV}^{-1}$ (extracted from $D(t)$).
• These values are in excellent agreement with independent dispersive determinations (Cao et al.), despite the methodological differences (lattice vs. scattering data).

D. Physical Pion Mass Prediction

Using the extracted LECs ($c_8, c_9$) and ChPT expressions, the authors evaluated the GFFs at the physical pion mass ($m_\pi \approx 139$ MeV).

• Nucleon D-term: $D(0) = -4.3 \pm 0.8$.
• This result is consistent with existing lattice, dispersive, and ChPT bounds.
• The framework also successfully reconstructed $D(t)$ using sparse experimental data from Deeply Virtual Compton Scattering (DVCS) when combined with the ChPT forward-limit constraint.

5. Significance and Outlook

• Paradigm Shift: This work demonstrates that generative AI, specifically diffusion models, can serve as a powerful tool for inverse problems in high-energy physics, replacing traditional parametric fitting with a flexible, data-driven prior.
• Efficiency: It proves that high-precision physics insights (like the D-term) can be extracted from very sparse datasets if the physical prior is sufficiently rich.
• Future Applications: The framework is extensible to other hadrons (e.g., $\Delta$-resonances, $\rho$-mesons), nuclear systems, and Generalized Parton Distributions (GPDs). It offers a pathway to constrain GPDs using experimental Compton Form Factors while satisfying fundamental constraints like polynomiality and positivity.
• Validation: The agreement between the diffusion-extracted LECs and independent dispersive methods provides a non-trivial cross-validation of both the machine learning approach and the underlying theoretical frameworks.

In summary, the paper establishes a new standard for reconstructing hadronic structure functions, combining the rigor of theoretical constraints with the flexibility of modern generative deep learning to overcome the limitations of sparse and noisy experimental/lattice data.