Unified Extraction of In-Medium Heavy Quark Potentials from RHIC to LHC Energies via Deep Learning

This paper employs deep learning within a Bayesian framework to simultaneously extract the real and imaginary components of the in-medium heavy quark potential from bottomonium suppression data across RHIC and LHC energies, revealing that while the real part remains close to the vacuum form, the imaginary part is the dominant driver of bottomonium suppression.

The Gist

The Big Picture: The "Ghostly" Soup and the Heavy Quark Twins

Imagine you smash two heavy atoms together at nearly the speed of light. For a tiny fraction of a second, the matter inside them melts into a super-hot, super-dense soup called Quark-Gluon Plasma (QGP). Scientists think this is what the universe looked like microseconds after the Big Bang.

Inside this soup, there are heavy particles called bottom quarks. When a bottom quark and its anti-particle (an anti-bottom) are created, they try to stick together to form a "twin" called Bottomonium.

The big mystery is: How does this hot soup affect these twins? Does the soup melt them apart? Does it change how they hold hands? To answer this, scientists need to know the "force field" (the potential) that exists inside the soup. But they can't see the soup directly; they can only see the twins after they escape.

The Problem: A Puzzle with Missing Pieces

Scientists have been trying to figure out the rules of this force field for decades. They have data from two giant particle smashers:

1. RHIC (in the US, lower energy).
2. LHC (in Europe, higher energy).

The problem is that the data is messy. Different theories give different answers. It's like trying to guess the shape of a hidden object by looking at its shadow from only one angle. You need to look at the shadows from many angles (different energies) to get the full picture.

The Solution: The "AI Detective"

This paper introduces a new way to solve the puzzle using Deep Learning (a type of Artificial Intelligence). Think of the researchers as detectives trying to find a specific criminal (the force field) based on the crime scene evidence (the experimental data).

Here is how they did it, step-by-step:

1. Building the "Simulation Factory"

Before they could use AI, they had to teach it what the world looks like.

• The Physics Engine: They built a complex computer simulation (based on quantum mechanics) that acts like a virtual laboratory.
• The Variables: They created thousands of "fake" universes. In each one, they tweaked the rules of the force field (making the soup hotter, the force stronger, or the distance different).
• The Output: For every set of rules, the simulation told them: "If the force field is this, the bottomonium twins will survive this much."

2. The Data Augmentation (The "Photocopier")

They didn't have enough fake universes to train a super-smart AI. So, they used a clever trick called Principal Component Analysis and Gaussian Process Regression.

• Analogy: Imagine you have 1,000 photos of a cat. You want to teach a computer to recognize cats. Instead of just showing it those 1,000 photos, you use a "smart photocopier" to generate 10,000 new, slightly different photos of cats (different angles, lighting, etc.) that still look like the original 1,000.
• This gave them a massive dataset to train their AI.

3. Training the Neural Network (The "Translator")

They built a Convolutional Neural Network (CNN).

• The Job: This AI acts as a translator.
  • Input: The rules of the force field (e.g., "The soup is this hot, the force is this strong").
  • Output: The predicted survival rate of the twins (the "shadow" or data).
• They trained the AI until it could perfectly predict the shadow based on the rules. It learned the complex, non-linear relationship between the soup and the twins.

4. The Reverse Engineering (The "Sherlock Holmes" Moment)

This is the most important part. Usually, you go from Rules $\to$ Data. But the scientists wanted to go from Data $\to$ Rules.

• They took the real experimental data from the LHC and RHIC (the actual shadows observed in the real world).
• They used a method called Stochastic Gradient Langevin Dynamics (SGLD).
• Analogy: Imagine you have a locked safe (the real data) and a machine that generates keys (the AI). You don't know the combination. You start turning the dial randomly. Every time you try a combination, the machine tells you how close you are to opening the safe. The AI helps you "feel" your way to the perfect combination that opens the safe for all the different experiments at once.

The Findings: What Did They Discover?

After running this massive AI reverse-engineering process, they found two surprising things about the "force field" inside the hot soup:

1. The "Real" Part (The Glue) is Weak:

   • Scientists thought the hot soup would act like a strong magnet that pulls the twins apart (color screening).
   • The Result: The AI found that the "glue" holding the twins together inside the soup is actually very similar to the glue in a vacuum (empty space). The soup doesn't weaken the glue as much as we thought. It's like putting a magnet in a hot room; the magnet still works almost the same way.

2. The "Imaginary" Part (The Friction) is Strong:

   • In quantum physics, "Imaginary" doesn't mean "fake." It represents dissipation or friction. It's the chance that the twins will crash into particles in the soup and get knocked apart.
   • The Result: This is the dominant factor! The reason the twins disappear isn't because the glue weakens; it's because they are constantly getting hit and scattered by the hot soup. It's like trying to walk through a crowded dance floor; you aren't falling because the floor is slippery (weak glue), but because people are bumping into you (friction/imaginary part).

Why This Matters

• Unification: They managed to find one single set of rules that explains the data from both the low-energy RHIC and the high-energy LHC. This is a huge step forward.
• The AI Advantage: Traditional methods struggled to handle the complexity of combining data from different energies. The AI was able to see patterns that human math couldn't easily find.
• The Conclusion: The hot soup doesn't melt the "glue" of the heavy quarks; it just knocks them around.

Summary in One Sentence

The researchers used a super-smart AI to reverse-engineer the rules of a subatomic "soup," discovering that the soup doesn't weaken the bond between heavy particles, but instead knocks them apart through constant collisions.

Technical Summary

1. Problem Statement

The study addresses the challenge of quantitatively determining the in-medium heavy quark (HQ) potential ($V(r, T)$) within the Quark-Gluon Plasma (QGP). While the suppression of heavy quarkonia (specifically bottomonium, $\Upsilon$) is a key signature of deconfinement, the precise nature of the potential governing the $b\bar{b}$ dipole evolution remains debated.

• Theoretical Gap: Lattice QCD calculations provide potentials at finite temperatures, but extraction methodologies vary, leading to significant uncertainties. Phenomenological models often rely on specific assumptions about color screening (real part) and inelastic scattering (imaginary part).
• Experimental Challenge: Experimental data exists across a wide range of collision energies (RHIC: 200 GeV Au-Au; LHC: 2.76 TeV and 5.02 TeV Pb-Pb). However, extracting a unified potential that simultaneously describes data from all these systems is difficult due to the non-linear, high-dimensional relationship between the potential parameters and the observed nuclear modification factors ($R_{AA}$).

2. Methodology

The authors propose a Deep Learning framework under a Bayesian perspective to solve this inverse problem. The methodology consists of four main stages:

A. Physical Modeling (Forward Problem)

• Schrödinger Equation: The evolution of the $b\bar{b}$ wave function is modeled using a time-dependent Schrödinger equation coupled with a hydrodynamic background (MUSIC package).
• Potential Parametrization: The complex in-medium potential $V(T, r) = V_R + iV_I$ is parameterized:
  • Real Part ($V_R$): Modeled using a screened Cornell potential with a Debye mass $\mu_D(T)$ dependent on a scaling parameter $a_4$.
  • Imaginary Part ($V_I$): Parameterized as $V_I = -iT a_0 (a_1 r + a_2 r^{a_3})$, capturing temperature and spatial dependencies.
  • Switching Temperature ($T_{sw}$): A parameter defining where the medium potential transitions to the vacuum Cornell potential.
• Hydrodynamics: The medium evolution is simulated using the MUSIC package with initial conditions derived from the Glauber model, covering 5.02 TeV, 2.76 TeV, and 200 GeV systems.

B. Dataset Generation and Augmentation

• Sampling: 1,000 random samples of the 6-dimensional parameter vector $\theta = (a_0, a_1, a_2, a_3, a_4, T_{sw})$ are generated.
• Simulation: For each sample, the Schrödinger equation is solved to generate theoretical $R_{AA}$ values for $\Upsilon(1S, 2S, 3S)$ across various centralities and $p_T$ bins.
• Augmentation: To overcome the scarcity of theoretical samples, Principal Component Analysis (PCA) is used for dimensionality reduction, followed by Gaussian Process Regression (GPR) to generate an augmented dataset of 10,000 events. This ensures the neural network covers the parameter space effectively.

C. Deep Learning Architecture (CNN)

• Model: Three independent Convolutional Neural Networks (CNNs) are trained, one for each collision energy (5.02 TeV, 2.76 TeV, 200 GeV).
• Mapping: The CNNs learn the non-linear mapping from the potential parameters (Input) to the multi-dimensional $R_{AA}$ observables (Output).
• Validation: The models are validated using a hold-out test set, showing excellent agreement between CNN predictions and the ground-truth Schrödinger model results ($R^2 > 0.96$).

D. Bayesian Inverse Extraction (SGLD)

• Objective: To find the potential parameters that best describe all experimental data simultaneously.
• Algorithm: Stochastic Gradient Langevin Dynamics (SGLD) is employed. This allows for sampling from the posterior distribution of the parameters rather than just finding a single point estimate.
• Loss Function: A joint loss function is constructed by summing the $\chi^2$ contributions from all three collision systems, weighted by experimental uncertainties.
• Process: The algorithm iteratively updates the parameters in an auxiliary unconstrained space to minimize the loss, generating a posterior distribution for the parameters.

3. Key Contributions

1. Unified Multi-Energy Extraction: Unlike previous studies that often analyzed energies separately, this work extracts a single, unified set of potential parameters that simultaneously describes bottomonium suppression from RHIC (200 GeV) to LHC (5.02 TeV).
2. Deep Learning for Inverse Problems: The paper demonstrates a robust pipeline using CNNs to emulate complex quantum mechanical simulations (Schrödinger + Hydro) and SGLD for Bayesian inference, significantly accelerating the extraction process compared to traditional Markov Chain Monte Carlo (MCMC) methods on the raw simulation.
3. Data Augmentation Strategy: The integration of PCA and GPR to expand the training dataset from 1,000 to 10,000 events effectively mitigates the "small data" problem inherent in computationally expensive physics simulations.
4. Validation via Mock Data: A rigorous "closure test" was performed using synthetic data generated from a known "ground truth" potential. The framework successfully recovered the input parameters, validating the reliability of the extraction pipeline.

4. Key Results

• Real Part of the Potential: The extracted real part ($V_R$) remains remarkably close to the vacuum Cornell potential. The Debye mass scaling parameter $a_4$ is found to be small ($\approx 0.1$), suggesting that color screening is weak in the temperature range accessible in these collisions. This contradicts some expectations of strong screening but aligns with recent lattice QCD and Bayesian analyses.
• Imaginary Part of the Potential: The imaginary part ($V_I$) is strongly constrained by the data and provides the dominant contribution to bottomonium suppression. The optimal parameters indicate a linear temperature dependence ($a_0 \approx 1$) and a spatial dependence dominated by terms up to $r^2$ ($a_3 \approx 2$).
• Switching Temperature: The extracted switching temperature $T_{sw}$ is consistent with the pseudo-critical temperature ($T_c \approx 0.17$ GeV), implying the current parametrization is robust and does not require an artificial hybrid transition region.
• Consistency: The study confirms that extracting potentials independently for each energy leads to discrepancies (particularly in the Debye mass parameter $a_4$), whereas the joint multi-energy extraction yields a consistent and physically robust potential.

5. Significance

This work represents a significant advancement in the field of heavy-ion physics by:

• Resolving Ambiguities: It provides strong evidence that the suppression of bottomonium from RHIC to LHC is driven primarily by the imaginary part of the potential (inelastic scattering/dissociation) rather than strong color screening of the real part.
• Methodological Benchmark: It establishes a new standard for using Deep Learning and Bayesian inference to constrain QCD properties from experimental data, offering a scalable framework for future studies of other observables (e.g., open heavy flavor, jet quenching).
• Theoretical Insight: The finding of a weakly screened real potential challenges some theoretical models and suggests that the in-medium heavy quark potential retains much of its vacuum structure even at high temperatures, with the imaginary component playing the critical role in dissociation.