Deep learning approaches to extract nuclear deformation parameters from initial-state information in heavy-ion collisions

This study demonstrates that deep learning models, particularly simulation-based inference with conditional normalizing flows, can effectively extract nuclear quadrupole and hexadecapole deformation parameters from initial-state configurations in heavy-ion collisions by leveraging multi-event averaging to suppress stochastic fluctuations and provide robust uncertainty quantification.

The Gist

Imagine trying to figure out the shape of a hidden object just by looking at the ripples it makes in a pond. That is essentially what physicists are doing when they smash heavy atoms together at near the speed of light. They want to know the shape of the atomic nuclei (the "hidden object") before the crash, but the collision is so chaotic that the signal gets lost in the noise.

This paper is about using Artificial Intelligence (AI) to act as a super-powered detective, helping scientists "see" the shape of these atomic nuclei by analyzing the chaos of the collision.

Here is a breakdown of the paper's journey, using simple analogies:

1. The Mystery: The Squishy, Wobbly Ball

Atomic nuclei aren't always perfect spheres. Some are like rugby balls (elongated), and some are like flattened pancakes (squashed). Physicists call these shapes "deformations."

• The Problem: When two nuclei collide, they don't just bounce; they explode into a hot soup of particles (quark-gluon plasma). The initial shape of the nuclei leaves a fingerprint on this explosion, but the explosion is also full of random "static" or noise. It's like trying to hear a whisper in a crowded, screaming stadium.

2. The Detective's Toolkit: Two Types of AI

The researchers tested two different AI strategies to decode this whisper:

• The "Point-Cloud" Detective (Microscopic View):

  • The Setup: Imagine looking at a bag of marbles. Each marble is a proton or neutron inside the nucleus.
  • The AI: They built an AI that looks at the 3D positions of every single marble.
  • The Trick: One marble is wobbly and noisy. But if you look at 20 bags of marbles all shaped the same way, the AI can average out the wobble and see the true shape.
  • The Result: This worked incredibly well. By grouping many "bags" (events) together, the AI could predict the shape with near-perfect accuracy. It proved that the shape information is there, hidden in the details.

• The "Entropy Map" Detective (Macroscopic View):

  • The Setup: In real life, we can't see the individual marbles. We only see the "heat map" or "smoke" left behind after the crash (called the entropy density profile). This is like looking at a blurry, pixelated photo of the explosion.
  • The AI: They used a more advanced AI (similar to the ones that recognize cats in photos) to look at these blurry heat maps.
  • The Challenge: A single blurry photo is almost impossible to read. The noise is too high.
  • The Solution: Just like with the marbles, they fed the AI many blurry photos at once (grouping 10, 50, or 100 collisions together). The AI learned to ignore the random static and focus on the common pattern.

3. The Two AI Personalities: The Guesser vs. The Statistician

The paper compared two ways the AI could answer the question "What is the shape?":

• The Guesser (Standard Regression):

  • This AI looks at the data and says, "I think the shape is a rugby ball tilted at 30 degrees." It gives you one single best guess.
  • Pros: It's fast and usually gets the center of the answer right.
  • Cons: It doesn't tell you how confident it is. It might be wildly wrong, but it won't tell you.

• The Statistician (Simulation-Based Inference - SBI):

  • This AI is more cautious. Instead of giving one number, it draws a map of possibilities. It says, "I think it's a rugby ball, but there's a 90% chance it's between 25 and 35 degrees, and a 10% chance it's a pancake."
  • Pros: It gives you the full picture of uncertainty. It tells you how sure it is.
  • Cons: It's computationally heavier, but the paper found it was actually better at finding the tricky shapes (the "hexadecapole" or diamond-like distortions) when enough data was provided.

4. The Big Discovery: "More is Better"

The most important lesson from this paper is the power of averaging.

• If you look at one collision, the AI is blind. The noise drowns out the shape.
• If you look at 100 collisions together, the noise cancels itself out, and the shape shines through clearly.
• The AI needs a "crowd" to hear the "whisper."

5. Why This Matters

This research is a crucial first step. It proves that the shape of an atomic nucleus is encoded in the very first moment of a collision.

• Before: Scientists had to guess the shape based on complex theories.
• Now: We have a blueprint showing that AI can extract this shape directly from the data.
• Future: Once this AI is trained on the "initial state" (the moment of impact), scientists can eventually use it to work backward from the final explosion (the particles flying out) to figure out the shape of nuclei we've never studied before.

In a nutshell: The paper shows that by using AI to look at many blurry photos of atomic collisions at once, we can finally "see" the hidden shapes of the atoms involved, turning a chaotic explosion into a clear picture of nuclear structure.

Technical Summary

1. Problem Statement

Heavy-ion collisions (e.g., at RHIC and LHC) create a Quark-Gluon Plasma (QGP), and the initial geometry of the colliding nuclei significantly influences final-state observables like anisotropic flow. A critical goal is to infer intrinsic nuclear structure parameters, specifically the quadrupole ($\beta_2$) and hexadecapole ($\beta_4$) deformation moments, from collision data.

However, this extraction is challenging due to:

• Event-by-event fluctuations: Stochastic variations in nucleon positions and entropy deposition obscure the underlying deformation signal.
• Non-linear correlations: The relationship between initial-state geometry and final observables is complex.
• Microscopic vs. Macroscopic: It is unclear how much information about deformation is preserved when moving from microscopic nucleon configurations to macroscopic entropy-density profiles (which are what experiments effectively probe via models like TRENTo).

The paper aims to determine the intrinsic identifiability of $\beta_2$ and $\beta_4$ from initial-state information using deep learning, establishing an upper bound on what can be extracted before hydrodynamic evolution further blurs these signals.

2. Methodology

The study employs a two-stage approach using deep learning architectures designed for set-based and image-based data.

A. Data Generation

1. Microscopic Level (Baseline): Nucleon configurations for $^{238}\text{U}$ are sampled from a deformed Woods–Saxon distribution defined by $\beta_2 \in [-0.5, 0.5]$ and $\beta_4 \in [-0.2, 0.2]$.
2. Macroscopic Level (Realistic): Initial entropy-density profiles are generated using the TRENTo model for $^{238}\text{U} + ^{238}\text{U}$ collisions at $\sqrt{s_{NN}} = 193$ GeV.
   • Inputs include 2D entropy density maps ($100 \times 100$) and 7 global attributes (impact parameter, eccentricities $e_2$–$e_5$, multiplicity, participant number).
   • Data is stratified by centrality bins (0–10%, 30–40%, 60–70%, 90–100%, and 0–100%).

B. Neural Network Architectures

The authors utilize multi-event grouping (bagging) to suppress stochastic fluctuations. A "bag" consists of $N$ (for nucleons) or $M$ (for entropy profiles) events sharing the same deformation parameters.

1. Point-Cloud Network (for Nucleon Configurations):

   • Architecture: A permutation-invariant network using PointMLP, Residual Blocks, and an Attention-based Multiple Instance Learning (MIL) mechanism.
   • Mechanism: It processes individual nucleon coordinates, aggregates features via max-pooling, and then uses two branches: an Attention branch (learning weights for specific configurations) and an Average branch (uniform averaging). These are concatenated to form a group-level representation.
   • Output: Regression head predicting $\beta_2$ and $\beta_4$.

2. Entropy-Density Network (for TRENTo Profiles):

   • Backbone: ConvNeXt-Small to encode 2D entropy maps into feature vectors.
   • Fusion: Global attributes are encoded via an MLP and concatenated with image features.
   • Aggregation: A Set Transformer (using Self-Attention Blocks and Pooling-by-Multihead-Attention) aggregates the variable-sized bag of events into a single representation.
   • Inference Strategies:
     • Standard Regression: A lightweight MLP head outputs point estimates of $\beta_2$ and $\beta_4$ (trained with Huber loss).
     • Simulation-Based Inference (SBI): Replaces the regression head with a Conditional Normalizing Flow (RealNVP). This models the full posterior distribution $p(\beta_2, \beta_4 | \text{data})$, providing calibrated uncertainty quantification.

3. Key Contributions

• Establishment of Intrinsic Identifiability: The study quantifies the theoretical upper limit of extracting deformation parameters from initial states, separating the signal from stochastic noise.
• Demonstration of Multi-Event Averaging: It proves that aggregating multiple events (increasing bag size $N$ or $M$) is essential to suppress fluctuations and reveal deformation signals, particularly for the more subtle $\beta_4$.
• Comparison of Regression vs. SBI: The paper provides a rigorous comparison showing that while standard regression captures central trends efficiently, SBI offers superior uncertainty quantification and robustness, especially for the hexadecapole parameter.
• Centrality Dependence Analysis: It systematically maps how inference accuracy degrades in peripheral collisions due to reduced participant numbers and weaker geometric imprints.

4. Key Results

A. Microscopic Level (Nucleon Configurations)

• Performance: As the group size $N$ increases from 1 to 20, the coefficient of determination ($R^2$) improves significantly.
  • At $N=20$, $R^2$ reaches >0.99 for $\beta_2$ and ~0.96 for $\beta_4$.
• Observation: $\beta_2$ is easier to infer than $\beta_4$ because quadrupole deformation dominates the global ellipticity, whereas $\beta_4$ affects higher-order surface curvature and is more easily masked by fluctuations.
• Conclusion: The network can effectively learn the underlying deformation if stochastic noise is averaged out.

B. Macroscopic Level (TRENTo Entropy Profiles)

• Regression Performance:
  • For single events ($M=1$), performance is poor ($R^2 \approx 0.25$ for $\beta_2$, $\approx 0.08$ for $\beta_4$).
  • With $M=100$, $R^2$ improves to ~0.93 for $\beta_2$ and ~0.62 for $\beta_4$.
  • Degeneracy: A characteristic "X-shape" pattern is observed where prolate ($\beta_2 > 0$) and oblate ($\beta_2 < 0$) shapes are hard to distinguish for small $M$, but this degeneracy resolves as $M$ increases.
• Simulation-Based Inference (SBI) Performance:
  • SBI achieves $R^2$ values comparable to or slightly better than regression (e.g., $R^2 \approx 0.94$ for $\beta_2$ at $M=100$).
  • Uncertainty Calibration: For $M=10$, the posterior calibration is nearly ideal. For larger $M$ ($50, 100$), the model becomes slightly conservative (over-coverage), producing wider credible intervals to account for the sharpness of the signal.
  • Posterior Geometry: As $M$ increases, the posterior distributions shrink from broad/distorted shapes to tight, Gaussian-like ellipses centered on the true values.

C. Centrality Dependence

• Inference quality systematically degrades in peripheral collisions (e.g., 90–100% centrality) where the number of participant nucleons is low, reducing the geometric imprint of deformation.
• Even with $M=100$, $R^2$ values drop significantly in peripheral bins compared to central collisions.

5. Significance and Outlook

• Foundational Bridge: This work establishes that nuclear deformation information is effectively encoded in the initial state and can be recovered with sufficient statistical averaging. This validates the feasibility of using heavy-ion collisions as a probe for nuclear structure.
• Methodological Advance: It demonstrates that Simulation-Based Inference (SBI) is a powerful tool for high-energy nuclear physics, offering a complete probabilistic characterization of parameters that standard regression cannot provide.
• Future Directions: The authors propose extending this framework to include full hydrodynamic evolution and final-state observables. Successfully identifying deformation in the initial state is a necessary prerequisite for extracting these parameters from experimental data, where viscous dissipation further blurs geometric features.

In summary, the paper confirms that deep learning, specifically when combined with multi-event averaging and SBI, can robustly extract nuclear deformation parameters from initial-state models, laying the groundwork for future experimental applications in nuclear structure physics.