Neural Unfolding of the Chiral Magnetic Effect in Heavy-Ion Collisions

TL;DR

This work tackles the CME detection challenge in heavy-ion collisions by training a time-embedded U-Net to time-reversely unfold CME-related charge separation from AMPT-based simulations. By using two-channel $p_T-\\phi$ distributions and iteratively predicting stepwise changes in the charge-separation field, the method reconstructs the full CME evolution across the QGP and hadronic phases. The results show high fidelity across multiple configurations, including robustness to varying initial CS strengths and inputs, indicating a promising route to infer early-time CME signals from late-stage observables. The approach offers a model-informed deep-learning framework that could enhance CME studies and motivate extensions incorporating chiral-kinetic dynamics for greater realism and experimental relevance.

Abstract

The search for the chiral magnetic effect (CME) in relativistic heavy-ion collisions (HICs) is challenged by significant background contamination. We present a novel deep learning approach based on a U-Net architecture to time-reversely unfold the dynamics of CME-related charge separation, enabling the reconstruction of the physics signal across the entire evolution of HICs. Trained on the events simulated by a multi-phase transport model with different cases of CME settings, our model learns to recover the charge separation based on final-state transverse momentum distributions at either the quark-gloun plasma freeze-out or hadronic freeze-out. This devises a methodological tool for the study of CME and underscores the promise of deep learning approaches in retrieving physics signals in HICs.

Figures (7)

• Figure 1: The evolution of the CME signal in the QGP medium, displayed in the transverse $X$-$Y$ plane at different times: $t = 0.2, 5.0, 10.0, 15.0$, and $20.0 \,\mathrm{fm}/c$. The upper panels show the non-interacting scenario ($\sigma = 0$ mb), while the lower panels present the case with partonic interactions ($\sigma = 3$ mb). The CME signal partons are indicated by triangular markers (two colors denotes positive and negative charges), with the background QGP partons represented by gray circles.
• Figure 2: Schematic illustration of the U-Net architecture used in our analysis. Each blue box corresponds to a multi-channel feature map, annotated with channel numbers (top) and spatial dimensions (bottom left). The white boxes represent copied feature maps. The arrows denote the different operations. The network processes two-channel input representing the transverse momentum distribution $(p_T, \phi)$ of final-state particles, with positive and negative charges encoded as separate channels (red and blue colormaps, respectively). The output predicts the difference between distributions at adjacent time steps.
• Figure 3: The inverse evolution of the charge separation (CS) fraction in the QGP phase. The star symbol (most right) indicates the input CS fraction at the freeze-out of QGP phase, while circular and square markers represent the ground truth and model predictions, respectively, at various early time points of QGP evolution. The gray arrows indicate the direction of the model's backward prediction, starting from the QGP freeze-out stage and tracing back to the initial QGP step by step in time.
• Figure 4: Prediction accuracy of event-average charge separation (CS) fraction at each time step for the three model configuration cases in Table \ref{['tab:model_configs']}, where blue, orange, and green symbols represent Case 1, Case 2, and Case 3, respectively.
• Figure 5: Two-dimensional histograms comparing model predictions with ground truth values for the charge separation(CS) fraction at selected time points: $t = 1.0, 5.0, 7.0, 9.0, 11.0, 13.0\,\mathrm{fm}/c$. The red dashed line ($y = x$) represents ideal agreement between predictions and ground truth values.