Unbiased Data-Driven Determination of the Nuclear Dipole Amplitude in the Color Glass Condensate
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: We're looking at the paper: "Unbiased Data-Driven Determination of the Nuclear Dipole Amplitude in the Color Glass Condensate".
Mira: Here's the gist of the paper. Gluon saturation limits the growth of parton densities at small Bjorken-x and is expected to be most pronounced in heavy nuclei The authors introduce a physics-informed neural-network framework that embeds the collinearly improved Balitsky-Kovchegov evolution equation directly into the training objective, allowing the impact-parameter-averaged dipole amplitude to be determined from data without assuming a functional form for its initial condition Applying this framework to forward-hadron nuclear-modification-factor and coherent J/psi photoproduction data, they extract the 208Pb dipole amplitude at x0 = 0.01 with QCD evolution and momentum-space positivity enforced throughout training Model Formulation The framework utilizes the collinearly improved Balitsky–Kovchegov (ciBK) equation to describe the energy evolution of the dipole amplitude, represented by a physics-informed neural network.
Kai: The evolution equation is given by Equation (2), which resums both double and single transverse logarithms using rapidity shifts and a collinearly improved kernel.
Mira: The network is trained to satisfy the ciBK residual while simultaneously describing the observable constraints.
Kai: Observable Constraints Two classes of data constrain the fit: 1.
Title and authors: Kai: So what does the paper actually claim they did in terms of their methodology? They are using a physics informed neural network to determine that impact parameter averaged dipole amplitude.
Mira: They embed the collinearly improved Balitsky-Kovchegov evolution equation directly into the training objective, which is the core idea here. It forces the network to obey the evolution laws while fitting observable data simultaneously.
Lev: So instead of just having a black box fit where we only check if it matches data, they are designing the learning process around satisfying two physics constraints at once.
Kai: And they constrain it using those forward hadron production and J/psi photoproduction datasets as the observables that guide the fitting.
Mira: They use a loss function L tot which combines violations of the ciBK evolution equation, constraints on nuclear modification factors, and positivity conditions in momentum space.
Lev: The paper mentions that they determined the weights for those different loss components using a gradient calibration procedure to make sure all these physics terms contribute at comparable scales to each other.
Kai: They also stated that solving the evolution equation directly inside a global fitting loop would be computationally too expensive, so they used this neural network representation as a way to get a smooth and differentiable function for the dipole amplitude N of r and Y.
Mira: That smoothness is key because it allows them to get good results at a substantially reduced evaluation cost compared to doing the direct numerical integration of that evolution equation.
Lev: So they are trading the complexity of solving differential equations directly for a neural network that learns the solution implicitly through training against physical laws and data.
Kai: That’s a big shift in how we approach these kinds of calculations, moving away from purely analytical or even parameterized starting points for the nuclear state.
Mira: It moves it toward using the available data to *determine* the physics of that initial state, rather than just fitting a model onto it.
Lev: If you were building real hardware for error correction, having an initial condition determined this way would give you a much more physically grounded starting point for your evolution simulations.
The paper's summary: Kai: They didn't just introduce the framework; they showed how it performs compared to the conventional methods we have used previously.
Mira: They showed that this data-driven approach is significantly better at handling the known model dependence, which was a major issue when people were constructing these initial conditions using things like simple scaling assumptions or eikonalizing the proton amplitude.
Lev: The paper notes that extracted values of the nuclear initial saturation scale and related quantities vary wildly in older literature, and reported values of Q two s zero Pb over Q two s zero p differ by factors of several depending on those underlying assumptions.
Kai: This paper tackles that directly by showing how their method avoids imposing a functional form on the nuclear initial condition entirely, which is what led to that wide variation in earlier results.
Mira: They also showed that when they extracted the Pb dipole amplitude at x zero point zero one, it reproduces both the coherent J/psi and RpPb data across the kinematic range they were testing.
Lev: That successful reproduction of two different types of observables is a strong validation that the framework is capturing something physically real about how these processes happen in lead.
Kai: They also quantified how much their saturation scale ratio Q two s zero Pb over Q two s zero p at x equal to zero point zero one actually aligns with simple geometric scaling, finding it to be three point one seven plus minus zero point one seven minus zero point one ten.
Mira: That comparison is a nice quantitative check because it shows that the dynamics they are modeling are leading them toward a result that matches simpler theoretical expectations in certain limits.
Lev: And for the initial condition itself, they found that it’s well described by an MV type form specifically for lead, which contrasts with the proton case, which is interesting because it reflects the higher color charge density of a large nucleus.
The paper's improvements: Kai: So to wrap up on this paper titled "Unbiased Data-Driven Determination of the Nuclear Dipole Amplitude in the Color Glass Condensate," they’ve shown a way to extract nuclear dipole amplitudes directly from data using this physics informed neural network approach.
Mira: The main implication is that we can determine these crucial initial conditions without having to rely on strong, hard-to-test assumptions about the nuclear profile, which was the biggest hurdle before.
Lev: For us in the error correction world, this means a much more physically grounded starting point for any evolution simulations we run, because you aren't starting with an arbitrary shape imposed by a theorist.
Kai: So we’re getting a determination of the 208Pb dipole amplitude at x zero point zero one that respects QCD evolution and momentum space positivity throughout the entire training process.
Mira: And they’ve successfully matched their result against both J/psi photoproduction and forward hadron production data, which gives confidence in the method's overall accuracy for these observables.
Lev: One thing to be careful about is that this framework relies on the chosen loss function, so if you change those physics terms, you might get a different result.
Kai: It’s definitely a dependence on the specific loss function they chose, but it’s an improvement over being completely blind to the nuclear structure.
Mira: In short, they’ve given us a new way to use data to constrain saturation physics by building in the evolution dynamics directly into the training process of an AI model.
Lev: The paper "Unbiased Data-Driven Determination of the Nuclear Dipole Amplitude in the Color Glass Condensate" gives us a tool to generate physically motivated initial conditions for nuclear structure studies at small x, which is a definite piece of utility for both theory and experiment.
Conclusion: Kai: So we just looked at "Unbiased Data-Driven Determination of the Nuclear Dipole Amplitude in the Color Glass Condensate," and what they did is extract that impact parameter averaged dipole amplitude directly from data without assuming any shape for the initial condition.
Mira: That’s right. They embedded the collinearly improved Balitsky Kovchegov evolution equation right into their training objective, which forces the AI to follow QCD rules while fitting those forward hadron production and J/psi photoproduction datasets simultaneously.
Lev: And they did this by using a loss function that includes terms for the evolution violations and positivity conditions in momentum space, making sure the initial condition is physically sound.
Kai: They showed that when you extract it for lead at x equal to zero point zero one, the resulting saturation scale ratio over proton is three point one seven plus minus zero point one seven minus zero point one ten, which matches simple geometric scaling quite well.
Mira: That’s a big deal because it means they can quantify how the nucleus behaves without having to rely on those older scaling assumptions that gave us so much variation in the literature.
Lev: From an error correction standpoint, knowing that the initial condition is determined by data and dynamics rather than a parameterized form gives us a much more realistic starting point for any evolution simulations we try to run on real hardware.
Kai: And they found that their method can predict the transverse momentum ratio in proton-proton, proton-lead, and lead-proton collisions without needing any system dependent parameters, matching those LHCb measurements at low multiplicity.
Mira: That’s strong evidence that the dynamics captured by this AI framework are describing the initial state evolution correctly.
Lev: They also found that their extracted initial condition for lead is well described by an MV type form, which is different from what they found for the proton, showing how much the color charge density in a large nucleus matters.
Kai: So yeah, to sum up "Unbiased Data-Driven Determination of the Nuclear Dipole Amplitude in the Color Glass Condensate," they’ve given us a data-driven way to find nuclear structure inputs that are consistent with QCD evolution.
Mira: It provides a phenomenological baseline for small x phenomenology at both the LHC and RHIC, and it allows us to interpolate this method to any arbitrary mass number A, giving us a new tool for studying nuclei beyond lead.
Lev: It stops the guesswork on those initial states and gives us something concrete to test when we get more small x sensitive observables from things like future electron ion colliders.
Kai: Next up, we’re going to look at how this impacts the search for dark matter particles through thermal evolution in hot nuclear matter.
Key Laboratory of Quark and Lepton Physics (MOE) & Institute of Particle Physics, Central China Normal University, Wuhan 430079, China · Artificial Intelligence and Computational Physics Research Center, Central China Normal University, Wuhan 430079, China · Institute of Frontier and Interdisciplinary Science, Key Laboratory of Particle Physics and Particle Irradiation (MOE), Shandong University, Qingdao, Shandong 266237
hep-ph, nucl-ex, nucl-th
Submitted: 2026-07-30
Updated: 2026-11-15
Comments: Updated the version that is accepted by the Physical Review D. The dipole amplitude table link is also attached
License: http://creativecommons.org/publicdomain/zero/1.0/
Importance score: 83/100
The gist: Gluon saturation limits the growth of parton densities at small Bjorken-x and is expected to be most pronounced in heavy nuclei The authors introduce a physics-informed neural-network framework that…
Key concepts
- Collinearly Improved Balitsky-Kovchegov (ciBK) Evolution
- This is a specific mathematical equation used to describe how the dipole amplitude evolves as the energy scale changes. It resums both double and single transverse logarithms, providing a more accurate description of gluon dynamics in the small-x regime within heavy nuclei.
- Physics-Informed Neural Network (PINN)
- A neural network trained not just on data, but also on physical laws. Here, the network is forced to satisfy the ciBK evolution equation during training. This ensures that the resulting dipole amplitude respects established QCD dynamics while fitting experimental observations.
- Dipole Amplitude
- This represents a fundamental quantity describing how color charges (partons) in a nucleus interact with each other at very small distances. Determining this amplitude allows physicists to probe the internal structure and saturation effects within heavy nuclei like lead.
- Saturation Scale Ratio ($Q_{s0,Pb}/Q_{s0,p}$)
- This ratio compares the characteristic momentum scale of gluon interactions in a lead nucleus versus a proton. The paper finds this ratio is consistent with simple geometric scaling, indicating how much the density of gluons is enhanced in the large nucleus compared to a single nucleon.
Terminology
Summary
Gluon saturation limits the growth of parton densities at small Bjorken-x and is expected to be most pronounced in heavy nuclei The authors introduce a physics-informed neural-network framework that embeds the collinearly improved Balitsky-Kovchegov evolution equation directly into the training objective, allowing the impact-parameter-averaged dipole amplitude to be determined from data without assuming a functional form for its initial condition Applying this framework to forward-hadron nuclear-modification-factor and coherent J/psi photoproduction data, they extract the 208Pb dipole amplitude at x0 = 0. p. 0 We introduce a physics-informed neural-network
framework that embeds the collinearly improved Balitsky-Kovchegov evolution equation directly into
the training objective, allowing the impact-parameter-averaged dipole amplitude to be determined
from data without assuming a functional form for its initial condition. Applying this framework to
forward-hadron nuclear-modification-factor and coherent J/ψ photoproduction data, we extract the
208Pb dipole amplitude at x0 = 0.Where the paper says it
Model Formulation
The framework utilizes the collinearly improved Balitsky–Kovchegov (ciBK) equation to describe the energy evolution of the dipole amplitude, represented by a physics-informed neural network. The evolution equation is given by Equation (2), which resums both double and single transverse logarithms using rapidity shifts and a collinearly improved kernel. p. 1 At next-to-leading order, the BK
kernel receives collinearly enhanced double-logarithmic
corrections that destabilize the evolution, driving the
amplitude negative or otherwise unphysical at moderate rapidities [57, 58]. We therefore work with the
collinearly improved BK (ciBK) equation, which resums
both the double and single transverse logarithms to all
orders [31, 57, 59–61]: The unknown amplitude is represented by a
neural network,
N(r, Y ) −→ Nθ(r, Y ), (10)
where θ denotes the trainable parameters. Rather
than solving the ciBK equation repeatedly inside the
fit, the network is trained to satisfy the ciBK residual while simultaneously describing the observable constraints.Where the paper says it
Where the paper says it
Observable Constraints
Two classes of data constrain the fit:
-
Forward hadron production in pPb collisions at the LHC, which constrains the momentum-space dipole distribution through a hybrid factorization formula. p. 2
Where the paper says it
Forward hadron production. Forward hadron production in pA collisions constrains the momentum-space dipole distribution through the hybrid factorization formula, after convolution with projectile parton distribution functions (PDFs) and fragmentation functions (FFs).
-
Exclusive J/psi photoproduction in Pb-Pb ultraperipheral collisions, where the coherent cross section is proportional to the square of the target gluon dipole amplitude at the relevant dipole size. p. 1
Where the paper says it
This eliminates a persistent source of bias in small-x phenomenology, where conclusions about saturation can depend as strongly on the assumed functional form as on the underlying physics. Coherent J/ψ photoproduction in ultraperipheral collisions (UPCs) is particularly well suited to provide such a constraint: a quasi-real photon fluctuates into a cc¯ pair that scatters coherently off the target nucleus, and the coherent cross section is proportional to the square of the target gluon dipole amplitude at the relevant dipole size [48–50].
Training Mechanism
The training loss function is formulated as Ltot = wciBKLciBK + wRpA LRpA + wVMLVM + wblackLblack + wposLpos + wγLγ, which encodes violations of the ciBK evolution equation, constraints on nuclear modification factors, and positivity conditions. The weights are determined through a gradient-calibration procedure to ensure all loss components contribute at comparable gradient scales. p. 5 (24)
The loss weights used in the subsequent training are assigned directly as
wa =
Gref
Ga + ϵG
, a ∈ I, (25)
where ϵG > 0 prevents numerical instabilities when a
gradient norm becomes very small. For ϵG ≪ Ga, this
prescription yields
waGa ≃ Gref, a ∈ I, (26)
so that the different loss components contribute at comparable gradient scales at the end of the calibration stage.Where the paper says it
Key Findings
The analysis yields several significant results:
-
The extracted Pb dipole amplitude reproduces the available coherent J/psi and RpPb data across the measured kinematic range. p. 1
Where the paper says it
Second, the BK equation is embedded directly in the PINN training objective rather than imposed a posteriori through a fitting template, so the extracted amplitude is constrained to satisfy QCD evolution throughout training. We show that the extracted Pb dipole amplitude reproduces the available coherent J/ψ and RpPb data across the measured kinematic range.
-
The saturation-scale ratio Q2s0,Pb/Q2s0,p = 3. p. 0
17+0.17−0.10 at xB = 0.01 is consistent with simple geometric scaling. p. 0Where the paper says it
01 with QCD evolution and momentum-space positivity enforced throughout training. The evolved amplitude reproduces the measured cross sections across the available kinematic range and yields a saturation-scale ratio Q 2 s0,Pb/Q2 s0,p = 3.
Where the paper says it
The evolved amplitude reproduces the measured cross sections across the available kinematic range and yields a saturation-scale ratio Q 2 s0,Pb/Q2 s0,p = 3. 10, consistent with simple geometric scaling.
-
The extracted initial condition of Pb is well described by an MV-type form, in contrast to the proton case, reflecting the higher color-charge density of a large nucleus. p. 0
Where the paper says it
10, consistent with simple geometric scaling. The extracted Pb initial condition is well described by a McLerranVenugopalan-type form, in contrast to the proton, reflecting the higher color-charge density of a large nucleus.
-
The ratio of mean transverse momentum in pp, pPb, and Pbp collisions agrees with recent LHCb measurements at low multiplicity without any system-dependent parameters. p. 9
Where the paper says it
First, evaluating RpPb with fragmentation-function sets not used in training leaves the result essentially unchanged, confirming that this observable isolates genuine nuclear dipole dynamics from fragmentation uncertainties, as intended; RpPb is also likewise largely insensitive to higher-order QCD effects, the fragmentation functions and to the choice of factorization scale. Second, the same amplitude predicts the rapidity dependence of the transversemomentum ratio in pp, pPb, and Pbp collisions without introducing any system-dependent parameters, and this prediction agrees with recent LHCb data [56] in the low-multiplicity regime, where initial-state dynamics are expected to dominate.
Generalization
The method is generalized to an arbitrary nuclear mass number A by constructing the nuclear dipole amplitude at the initial rapidity corresponding to xB = 0.01 and evolving it in xB with the ciBK equation. p. 1 At next-to-leading order, the BK
kernel receives collinearly enhanced double-logarithmic
corrections that destabilize the evolution, driving the
amplitude negative or otherwise unphysical at moderate rapidities [57, 58]. We therefore work with the
collinearly improved BK (ciBK) equation, which resums
both the double and single transverse logarithms to all
orders [31, 57, 59–61]: Starting
from the unbiased, data-driven initial amplitudes for
the proton and 208Pb, we construct intermediate nuclei through the opacity interpolation described above,
and then evolve all systems in xB using the ciBK equation. These impact-parameter averaged initial conditions therefore provide benchmark inputs for small-x phenomenology at the LHC and RHIC, and for future measurements at the Electron–Ion Collider. The growing deviation at
higher multiplicities is a natural signature of final-state
effects that lie outside the scope of the present dipole-only
framework. Finally, we constructed physically motivated
initial conditions for nuclei intermediate between the proton and 208Pb by interpolating the extracted opacity,
providing a phenomenological baseline that can be tested
directly once additional small-x-sensitive observables become available for other nuclear species.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Physics-informed neural network framework embeds collinearly improved Balitsky-Kovchegov evolution equation directly into training objective, allowing impact-parameter-averaged dipole amplitude to be determined from data without assuming a functional form for its initial condition. This allows the system to extract nuclear structure based on data and dynamics rather than parametrized initial conditions.
-
The system can determine the 208Pb dipole amplitude at x0 = 0.01 with QCD evolution and momentum-space positivity enforced throughout training by minimizing a loss function that includes terms like LciBK, Lpos, and Lphys. This ensures the extracted amplitude satisfies QCD evolution while remaining positive in momentum space.
-
The improved AI system can predict the rapidity dependence of the transversemomentum ratio in pp, pPb, and Pbp collisions without any system-dependent parameters by using the extracted amplitude to make predictions that agree with LHCb measurements at low multiplicity. This provides a predictive capability grounded in initial-state dynamics.
-
The system can provide a quantitative test of geometric scaling by extracting the saturation-scale ratio Q2s0,Pb/Q2s0,p = 3.17+0.17−0.10 at xB = 0.01 and an effective transverse radius RPb = 7.83+0.15−0.12 fm for A = 208, which compares to the simple geometric expectation Q2s0,A/Q2s0,p approximately by about 20 percent.
-
The system can interpolate the extracted opacity to construct physically motivated initial conditions for nuclei intermediate between the proton and 208Pb, providing a phenomenological baseline that can be tested once additional small-x-sensitive observables become available for other nuclear species. This extends the methodology to arbitrary nuclear mass numbers A.
Abstract
Gluon saturation limits the growth of parton densities at small Bjorken- x and is expected to be most pronounced in heavy nuclei. Yet quantitative extractions of the nuclear gluon dipole amplitude have long relied on parametrized initial conditions, introducing uncontrolled model dependence that obscures genuine nuclear effects. We introduce a physics-informed neural-network framework that embeds the collinearly improved Balitsky-Kovchegov evolution equation directly into the training objective, allowing the impact-parameter-averaged dipole amplitude to be determined from data without assuming a functional form for its initial condition. Applying this framework to forward-hadron nuclear-modification-factor and coherent J/ψ photoproduction data, we extract the 208 Pb dipole amplitude at x 0=0.01 with QCD evolution and momentum-space positivity enforced throughout training. The evolved amplitude reproduces the measured cross sections across the available kinematic range and yields a saturation-scale ratio Q s0, Pb 2/Q s0,p squared = 3.17+0.17-0.10, consistent with simple geometric scaling. The extracted Pb initial condition is well described by a McLerran-Venugopalan-type form, in contrast to the proton, reflecting the higher color-charge density of a large nucleus. Using the same amplitude, we predict the rapidity dependence of the transverse-momentum ratio in pp, p Pb, and Pb p collisions, finding agreement with recent LHCb measurements at low multiplicity without any system-dependent parameters. This work provides the first unbiased, data-driven determination of nuclear structure in the saturation regime and establishes a general strategy for embedding nonlinear evolution equations into machine-learning extractions of dynamically constrained observables.
Sources
- The Color Glass Condensate and High Energy Scattering in QCD
- Evolution at small x_bj: The Color Glass Condensate
- The Color Glass Condensate
- Gluon saturation and initial conditions for relativistic heavy ion collisions
- Mining for Gluon Saturation at Colliders
- Computing Quark and Gluon Distribution Functions for Very Large Nuclei
- Gluon Distribution Functions for Very Large Nuclei at Small Transverse Momentum
- Green's Functions in the Color Field of a Large Nucleus
- Science Requirements and Detector Concepts for the Electron-Ion Collider: EIC Yellow Report
- The Electron-Ion Collider: Assessing the Energy Dependence of Key Measurements
- Electron Ion Collider: The Next QCD Frontier - Understanding the glue that binds us all
- Imaging the nucleus with high-energy photons
- Electron-Ion Collider in China
- Nuclear Cold QCD: Review and Future Strategy
- Operator expansion for high-energy scattering
- Small-x F_2 Structure Function of a Nucleus Including Multiple Pomeron Exchanges
- The Intrinsic Glue Distribution at Very Small x
- The BFKL Equation from the Wilson Renormalization Group
- The Wilson renormalization group for low x physics: towards the high density regime
Related papers
- MadNIS at NLO
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses