Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An End-to-End Generative Diffusion Model for Heavy-Ion Collisions

Jing-An Sun, Li Yan, Charles Gale, Sangyong Jeon

TL;DR

The paper tackles the computational bottleneck in high-statistics ultra-relativistic heavy-ion collision simulations. It introduces DiffHIC, an end-to-end conditional diffusion model that maps initial entropy density profiles $\mathbf{I}$ and the ratio $\eta/s$ to final-state spectra $\mathbf{S}$. DiffHIC delivers about $10^5$-fold speedup over traditional hybrid models while reproducing observables including integrated and differential anisotropic flow, multiparticle cumulants, and momentum fluctuations (e.g., $v_n$, $v_n\{2\}$, $v_n\{4\}$). The approach enables rapid parameter exploration and can be extended to 3D hydrodynamics, bulk viscosity, and full particle clouds, with quick fine-tuning for new physics such as nuclear deformation.

Abstract

Heavy-ion collision physics has entered the high precision era, demanding theoretical models capable of generating huge statistics to compare with experimental data. However, traditional hybrid models, which combine hydrodynamics and hadronic transport, are computationally intensive, creating a significant bottleneck. In this work, we introduce DiffHIC, an end-to-end generative diffusion model, to emulate ultra-relativistic heavy-ion collisions. The model takes initial entropy density profiles and transport coefficients as input and directly generates two-dimensional final-state particle spectra. Our results demonstrate that DiffHIC achieves a computational speedup of approximately $10^5$ against traditional simulations, while accurately reproducing a wide range of physical observables, including integrated and differential anisotropic flow, multi-particle correlations, and momentum fluctuations. This framework provides a powerful and efficient tool for phenomenological studies in the high-precision era of heavy-ion physics.

An End-to-End Generative Diffusion Model for Heavy-Ion Collisions

TL;DR

The paper tackles the computational bottleneck in high-statistics ultra-relativistic heavy-ion collision simulations. It introduces DiffHIC, an end-to-end conditional diffusion model that maps initial entropy density profiles and the ratio to final-state spectra . DiffHIC delivers about -fold speedup over traditional hybrid models while reproducing observables including integrated and differential anisotropic flow, multiparticle cumulants, and momentum fluctuations (e.g., , , ). The approach enables rapid parameter exploration and can be extended to 3D hydrodynamics, bulk viscosity, and full particle clouds, with quick fine-tuning for new physics such as nuclear deformation.

Abstract

Heavy-ion collision physics has entered the high precision era, demanding theoretical models capable of generating huge statistics to compare with experimental data. However, traditional hybrid models, which combine hydrodynamics and hadronic transport, are computationally intensive, creating a significant bottleneck. In this work, we introduce DiffHIC, an end-to-end generative diffusion model, to emulate ultra-relativistic heavy-ion collisions. The model takes initial entropy density profiles and transport coefficients as input and directly generates two-dimensional final-state particle spectra. Our results demonstrate that DiffHIC achieves a computational speedup of approximately against traditional simulations, while accurately reproducing a wide range of physical observables, including integrated and differential anisotropic flow, multi-particle correlations, and momentum fluctuations. This framework provides a powerful and efficient tool for phenomenological studies in the high-precision era of heavy-ion physics.
Paper Structure (8 sections, 2 figures)

This paper contains 8 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. The Generative Diffusion Model
  4. Training and Validation Datasets
  5. Results and Performance
  6. Computational Efficiency
  7. Physics Fidelity
  8. Conclusion and Outlook

Figures (2)

  • Figure 1: Centrality dependence of integrated anisotropic flow. Filled symbols represent the ground truth from the hybrid model, while colored bands show the results from DiffHIC. The agreement is excellent across different orders ($n$), cumulants, and shear viscosities.
  • Figure 2: Symbols represent results from a hydrodynamic model, while lines show the output of a Generative AI model. The top row shows the pre-trained model, and the bottom row shows the improved results after fine-tuning with 500 new events.