⚛️ phenomenology

Parton Fragmentation Functions Extracted with a Physics-Informed Neural Network

This paper introduces a novel Physics-Informed Neural Network (PINN) approach that eliminates the need for traditional parameterized forms and separate DGLAP evolution to directly extract non-perturbative parton fragmentation functions from experimental data, demonstrating their accuracy and universal applicability across a wide range of collision energies.

Original authors: Si-Wei Dai, Fu-Peng Li, Long-Gang Pang, Xin-Nian Wang, Ben-Wei Zhang, Han-Zhong Zhang

Published 2026-01-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Si-Wei Dai, Fu-Peng Li, Long-Gang Pang, Xin-Nian Wang, Ben-Wei Zhang, Han-Zhong Zhang

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 you are trying to understand how a giant, invisible cloud of energy (created when two particles smash together) turns into a shower of specific, tangible objects like marbles, beads, or pebbles. In the world of high-energy physics, this "cloud" is made of partons (tiny pieces of protons and neutrons), and the "objects" are hadrons (particles like protons and pions).

The rules for how this transformation happens are called Fragmentation Functions (FFs). Think of FFs as a "recipe book" that tells physicists exactly how likely a specific type of energy piece is to turn into a specific type of particle.

For a long time, scientists have tried to write down this recipe book by guessing the shape of the recipes (using mathematical formulas) and then tweaking them until they matched experimental data. This is like trying to guess a cake recipe by tasting the cake and adjusting the flour and sugar amounts blindly. It works, but it's slow, and you might miss the secret ingredient because you were stuck on a specific guess.

Here is what this new paper does:

1. The New Approach: The "Smart Chef" (PINN)

Instead of guessing a recipe shape, the authors built a Physics-Informed Neural Network (PINN). Imagine a super-smart chef who doesn't just taste the cake but also knows the laws of baking by heart.

No Guessing the Shape: Unlike old methods, this AI doesn't start with a pre-written formula. It learns the recipe directly from the data.
The "Law of Physics" Constraint: The most important part is that the AI is forced to follow the "laws of baking" (specifically, the DGLAP equations). These laws describe how the recipe changes as the "oven temperature" (energy level) changes.
- Analogy: If you bake a cake at a low temperature, it looks one way. If you bake it at a high temperature, it looks different. The old methods had to manually calculate these changes. This new AI has the rule "If temperature goes up, the cake changes this way" built directly into its brain. It doesn't need to be told to calculate the change; it just knows it must happen.

2. The Training: Learning from the "Taste Tests"

The team fed this AI data from electron-positron annihilation experiments.

Analogy: Imagine smashing a positive and negative electron together. They vanish and create a burst of energy that turns into a spray of particles. Scientists have measured exactly what comes out of this spray for decades.
The AI looked at these "sprays" (data) and adjusted its internal recipe book until the predictions matched the real-world observations perfectly.
They used a special mathematical trick called the Mellin Transform. Think of this as translating the recipe from "ingredients and steps" into a "code language" that the AI can solve much faster and more accurately, like translating a complex novel into a simple list of keywords to find the plot quickly.

3. The Results: A Better Recipe Book

The authors created a new set of Fragmentation Functions (the recipe book) using this method. Here is what they found:

It Works Everywhere: They tested their new recipe book not just on the electron collisions they trained it on, but also on proton-proton collisions (like those at the Large Hadron Collider).
- Analogy: It's like training a chef on making cakes in a small kitchen, then sending them to a massive industrial bakery. The chef still makes perfect cakes. The paper shows their AI-predicted recipes work perfectly across a huge range of energy levels, from low-energy collisions to the highest energies humans have ever created.
Better at the Edges: The new recipes seem to handle the "edges" of the data (where particles carry almost all the energy or almost none) better than previous methods.
The Glue Problem: The paper admits that because they only used data from electron collisions, they aren't 100% sure about the "glue" part of the recipe (the gluon fragmentation).
- Analogy: Electron collisions are great for seeing how the "flour" (quarks) behaves, but they don't show the "yeast" (gluons) as clearly. The authors note that if they add data from proton collisions (which are better at showing the yeast), their recipe for the glue will become even more precise.

4. Why This Matters

The paper claims this method is a significant step forward because:

It's Faster: It removes the need for slow, manual calculations of how the recipe changes with energy.
It's Unbiased: It doesn't force the data into a pre-existing mathematical box. It lets the data speak for itself while still obeying the laws of physics.
It's Universal: The resulting "recipe book" works reliably across different types of particle collisions and energy scales.

In short, the authors built a smart AI that learned the "laws of particle cooking" directly from the data, creating a more accurate and flexible guide for how energy turns into matter, without needing to guess the shape of the rules beforehand.

Technical Summary: Parton Fragmentation Functions Extracted with a Physics-Informed Neural Network

Problem Statement
Reliable predictions for high-energy strong interaction processes depend critically on non-perturbative Parton Fragmentation Functions (FFs). Traditionally, extracting FFs from experimental data requires assuming specific parameterized functional forms at an input scale and subsequently evolving them using the Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) evolution equations. This conventional approach introduces potential parametrization bias and relies on rigid functional forms that may not fully capture non-perturbative dynamics. Furthermore, the iterative numerical evaluation of DGLAP evolution and cross-section convolutions during global fits is computationally demanding, particularly when scanning large parameter spaces.

Methodology
The authors propose a novel approach utilizing a Physics-Informed Neural Network (PINN) to determine parton FFs without prior assumptions of parameterized forms. The methodology integrates the DGLAP evolution equations directly into the neural network architecture.

Network Architecture: The framework employs a cascaded design with two fully connected networks (NN1 and NN2). NN1 encodes the $Q^2$ -dependence via the evolution variable $\ln(Q^2)$ , while NN2 processes the combined input of the kinematic variable $z$ and the latent features from NN1 to output auxiliary functions. The FFs are constructed as $zD_i^{h\pm}(z, Q) = (1-z)^{\nu_i} \nu_i(z, Q)$ to enforce the boundary condition $D(z=1)=0$ and ensure positivity.
Physics Constraints: The network is trained using a composite loss function that minimizes deviations from experimental data while simultaneously penalizing violations of the DGLAP evolution equations. To improve numerical efficiency, the evolution constraints are enforced in Mellin space, where convolution integrals reduce to algebraic products.
Differentiation: Automatic differentiation is utilized within the optimization loop to compute the scale derivatives required for the DGLAP residuals, ensuring analytical precision without symbolic differentiation.
Data and Uncertainty: The extraction utilizes single-inclusive electron-positron annihilation (SIA) data from ALEPH, TASSO, TPC, OPAL, DELPHI, and SLD collaborations. The dataset includes inclusive, flavor-tagged (uds, c, b), and longitudinal cross-section measurements to constrain light, heavy, and gluon fragmentation. Experimental uncertainties are propagated using the Monte Carlo replica method, generating an ensemble of 100 pseudodata sets.

Key Contributions

Non-Parametric Extraction: The study demonstrates a method to extract FFs that does not rely on predefined functional forms, thereby reducing parametrization bias and allowing the data to dictate the functional shape.
Integrated Evolution: By embedding the DGLAP equations directly into the loss function, the extracted FFs automatically satisfy QCD evolution constraints across the $(z, \mu)$ domain, eliminating the need for separate evolution steps during the fitting process.
Mellin Space Implementation: The work successfully implements the inverse Mellin transform and DGLAP evolution within the neural network framework, leveraging the efficiency of Mellin moments for solving the coupled differential-integral equations.

Results

DGLAP Verification: The extracted FFs were verified against standard numerical DGLAP evolution. The PINN results at scales of 14 GeV and 1 TeV (the latter being outside the training data range) agreed with direct numerical evolution within 5% accuracy, confirming the network's ability to satisfy evolution equations.
Comparison with Existing Fits: The extracted FFs were compared with the AKK08 and KRE parameterizations. While the singlet quark FFs showed agreement in the range $z \in [0.05, 0.8]$ , discrepancies were observed for gluon and heavy quark FFs, particularly at large $z$ . The authors attribute the gluon discrepancy to the exclusive use of SIA data in this study, whereas AKK08 and KRE incorporated proton-proton collision data which provides stronger constraints on gluon fragmentation.
Closure Test: A closure test was performed where the extracted FFs were used to generate theoretical cross-sections, which were then used to retrain the network. The retrained results remained consistent with the original extraction (within a 20% deviation in the $z \in [0.01, 0.8]$ range), demonstrating the stability and reproducibility of the method.
Phenomenological Application: The extracted FFs were applied to calculate single-inclusive charged hadron spectra in proton-(anti)proton collisions at RHIC and LHC energies ( $\sqrt{s} = 130$ to $7000$ GeV). The Next-to-Leading Order (NLO) pQCD calculations using the PINN-extracted FFs aligned well with experimental data across this wide energy range, validating the universality of the extracted functions.

Significance
The paper claims that the PINN method represents a significant step forward in the extraction of non-perturbative quantities. By eliminating the need for parameterized forms and decoupling the fitting process from separate DGLAP evolution steps, the approach simplifies the extraction procedure and enhances the universal applicability of FFs across different energy scales. The authors posit that this framework offers a more robust and unbiased estimation of FFs, potentially uncovering features missed by traditional parametric methods, and lays the groundwork for future applications involving modified evolution equations in heavy-ion collisions.

1. The New Approach: The "Smart Chef" (PINN)

2. The Training: Learning from the "Taste Tests"

3. The Results: A Better Recipe Book

4. Why This Matters

Technical Summary: Parton Fragmentation Functions Extracted with a Physics-Informed Neural Network

More like this