Determination of the HERA coherent diffractive $J/\psi$ production cross section via artificial neural network

This paper presents a model-independent analysis of HERA's exclusive coherent diffractive $J/\psi$ production data using artificial neural networks to predict differential cross-sections and extract a $Q^2$- and $W$-dependent exponential slope by integrating HERA and LHC datasets.

The Gist

Imagine you are trying to understand the shape of a ghost. You can't see the ghost directly, but you can throw tiny, invisible ping-pong balls at it and watch how they bounce off. By studying the pattern of the bounces, you can figure out if the ghost is round, flat, or lumpy.

In the world of high-energy physics, scientists do something similar. They smash particles together to learn about the "shape" of protons (the building blocks of matter). Specifically, they look at a process where a photon (a particle of light) hits a proton and creates a heavy particle called a J/ψ meson, leaving the proton intact. This is like throwing a ball at a wall and having a new, heavy ball pop out while the wall stays standing.

Here is a simple breakdown of what this paper does, using everyday analogies:

1. The Old Way: Guessing with a Blueprint

For a long time, scientists tried to predict how these particles would bounce off each other using complex mathematical "blueprints" (theoretical models). These blueprints relied on many assumptions about how the proton looks inside and how the particles interact.

• The Problem: These blueprints were like trying to draw a map of a city using only a few street signs. They worked well in some neighborhoods (specific energy ranges) but got messy and unreliable in others. If the assumptions in the blueprint were slightly wrong, the whole map was wrong.

2. The New Way: The "Smart Learner" (Artificial Neural Network)

Instead of using a pre-drawn blueprint, the authors of this paper taught a computer an Artificial Neural Network (ANN)—essentially a digital brain—to learn the rules directly from the data.

• The Analogy: Imagine you have a huge photo album of every time someone threw a ball at a wall in the past (data from the HERA experiment). Instead of writing a rulebook on how the ball should bounce, you show the photos to a smart student. The student looks at thousands of examples and learns the patterns on their own: "Oh, when the ball is thrown harder, it bounces differently. When the wall is hit at a specific angle, the bounce changes."
• The Advantage: This "student" doesn't need to know the complex physics theory behind why the bounce happens. It just learns how it happens based on the evidence. This removes the bias of guessing the wrong blueprint.

3. The Training Process: The "Deep Ensemble"

To make sure their "student" wasn't just memorizing the answers or getting lucky, the scientists didn't just train one brain; they trained 100 different brains (a "Deep Ensemble").

• The Analogy: Imagine asking 100 different experts to look at the same photo album and guess the next bounce. If all 100 experts agree, you can be very confident in the answer. If they disagree, you know there is uncertainty.
• The Result: By averaging the answers of these 100 models, the scientists got a very reliable prediction that accounts for both the noise in the data and the uncertainty in the model itself.

4. What They Found

Using this "smart learner" approach, the team successfully predicted how the particles behave across a wide range of energies and angles, covering data from the HERA experiment and extending it to the LHC (Large Hadron Collider).

• The "Slope" Discovery: One key thing they measured was the "exponential slope" (a number called b). Think of this as measuring how "steep" the bounce is.
  • They found that this steepness isn't constant; it changes depending on how hard the photon hits (energy) and the type of collision.
  • Their "smart learner" confirmed that this slope depends heavily on the energy and the "virtuality" (how much energy the photon carries), matching what other experiments had seen but without needing the complex theoretical assumptions.

5. The Bottom Line

This paper shows that you don't always need a perfect theoretical theory to understand complex physics data. By using a data-driven approach (teaching a computer to learn from the data itself), they created a flexible tool that:

1. Avoids Guesswork: It doesn't rely on shaky assumptions about the proton's internal structure.
2. Handles Complexity: It can navigate the messy, multi-dimensional relationships between energy, angles, and particle types better than old methods.
3. Provides Confidence: It tells scientists not just the answer, but how sure they can be about that answer.

In short, the authors built a digital "pattern recognizer" that successfully mapped out the behavior of J/ψ particle production, proving that sometimes, letting the data speak for itself is the best way to understand the universe.

Technical Summary

Technical Summary: Determination of the HERA Coherent Diffractive J/ψ Production Cross Section via Artificial Neural Network

Problem Statement
Exclusive coherent diffractive $J/\psi$ production is a critical probe for understanding high-energy physics phenomenology, particularly the proton's gluon distribution and saturation effects. While extensive experimental data exists from the HERA collider (H1 and ZEUS collaborations) and the LHC, traditional theoretical analyses rely heavily on the "dipole picture" approach. This approach involves significant model dependencies, including assumptions about the vector meson wave function, the target profile, skewness corrections, and the real-imaginary part of the scattering amplitude. Furthermore, these theoretical models are often constrained to narrow kinematic ranges (typically small $|t|$ and moderate $Q^2$), limiting their predictive power across the full spectrum of available data. The authors identify a need for a model-independent approach that can handle multi-dimensional correlations in the data without relying on specific theoretical assumptions.

Methodology
The authors propose a data-driven framework utilizing Artificial Neural Networks (ANNs) to model the differential cross-section ($d\sigma/dt$) for exclusive coherent diffractive $J/\psi$ production. The methodology involves the following key components:

• Dataset: The model is trained on a combined dataset of 108 data points from the H1 and ZEUS collaborations at HERA. The kinematic coverage includes photon virtuality $0.05 < Q^2 < 100 \text{ GeV}^2$, photon-proton center-of-mass energy $20 < W < 250 \text{ GeV}$, and momentum transfer $|t| < 1.2 \text{ GeV}^2$.
• Input Features: Four variables are used: $Q^2$, $W$, $t$, and inelasticity $y$. To improve numerical stability, $Q^2$, $W$, and $y$ are transformed to a logarithmic scale, while $t$ remains linear.
• Network Architecture: The authors employ a "Deep Ensemble" approach consisting of $N=100$ architecturally identical models. Each model features three hidden layers (64, 64, and 32 neurons) with $tanh$ activation and L2 regularization.
• Uncertainty Quantification: The architecture is designed for heteroscedastic regression, outputting both the mean prediction ($\mu$) and the aleatoric uncertainty (variance). To address epistemic uncertainty (model sensitivity to initialization), the ensemble method aggregates predictions from 100 models trained with different random seeds. The final uncertainty combines both aleatoric (data noise) and epistemic (model variance) components.
• Loss Function: Training utilizes a Gaussian Negative Log Likelihood (NLL) loss function, which accounts for both experimental errors and the learned model uncertainty.
• Validation: The ensemble is evaluated using $\chi^2/\text{ndf}$ and Pull distributions on a held-out test set (10% of data).

Key Contributions

1. Model-Independent Prediction: The paper presents a fully data-driven approach that eliminates reliance on specific theoretical ingredients (e.g., dipole amplitudes or proton profiles) to predict cross-sections.
2. Robust Uncertainty Estimation: By combining Deep Ensemble methods with heteroscedastic regression, the framework provides a rigorous quantification of both experimental and model-induced uncertainties.
3. Extraction of the Exponential Slope ($b$): The differentiability of the ANN allows for the direct extraction of the exponential slope parameter $b$ (where $d\sigma/dt \propto \exp(-b|t|)$) across various kinematic regions, a task that is difficult with tree-based machine learning methods.
4. Extension to LHC Kinematics: The trained model is extended to predict total photoproduction cross-sections at higher energies (LHC range) by integrating the differential cross-section, despite the training data being limited to HERA energies.

Results

• Differential Cross-Section ($d\sigma/dt$): The ANN predictions show good agreement with HERA data across a broad range of $Q^2$ and $t$. The model successfully captures the decrease of the cross-section as $t$ and $Q^2$ increase. Minor underestimations were noted for specific $Q^2$ values at low $t$, and some discrepancies appeared at very low $W$ for photoproduction, likely due to large experimental uncertainties in those specific data points.
• Total Cross-Section ($\sigma$): The model reproduces the $W$-dependence of the total cross-section well within the HERA range. When extrapolated to LHC energies ($W > 250 \text{ GeV}$), the model underestimates data points from ALICE and LHCb, and exhibits large uncertainties, reflecting the lack of training data in this high-energy regime.
• Exponential Slope ($b$): The extracted slope parameter $b$ demonstrates a strong dependence on both $Q^2$ and $W$. The results are generally consistent with experimental data from H1 and ZEUS, yielding a mean value of $b = 4.72 \pm 0.15 \text{ (stat.)} \pm 0.12 \text{ (syst.)} \text{ GeV}^{-2}$ in the range $2 < Q^2 < 100 \text{ GeV}^2$.
• Model Performance: The ensemble achieved an average $\chi^2/\text{ndf} = 0.86 \pm 0.08$ and a Pull distribution consistent with a standard Gaussian ($\mu_{\text{pull}} \approx -0.10$, $\sigma_{\text{pull}} \approx 0.92$), indicating the model is unbiased and the uncertainty estimates are reliable.

Significance and Claims
The authors claim that this work demonstrates the viability of data-driven ANN frameworks as a complementary tool to traditional QCD-based models. By minimizing sensitivity to specific theoretical assumptions, the ANN approach offers a robust method for interpolating across multi-dimensional kinematic regions. The paper modestly concludes that while the method successfully captures nonlinear correlations and provides reliable uncertainty estimates, it is not a replacement for theoretical understanding but rather a tool to reduce model bias. The authors suggest that future work could involve hybrid frameworks (Physics-Informed Neural Networks) or the inclusion of additional vector meson data to further investigate small-$x$ saturation dynamics and the proton's transverse profile.