Harnessing data-driven methods for precise model independent event shape estimation in relativistic heavy-ion collisions

TL;DR

This work tackles the problem of classifying event topology in relativistic heavy-ion collisions through data-driven regression that predicts the observables $S_{0}$ and $S_{0}^{p_{ m T}=1}$ from global event inputs. By benchmarking a broad set of regression algorithms (including PR, DTR, ETR, KNN, LGBM, and MLP) and performing extensive hyperparameter optimization, the study identifies unweighted spherocity as the more predictive target and finds Light Gradient Boosting Machine to deliver the best overall accuracy and efficiency. SHAP analyses provide interpretability, showing that $\langle dN_{\rm ch}/d\eta \rangle$ dominates the predictions, with $\langle v_2 \rangle$ and $\langle m_T \rangle$ contributing less, while the reaction-plane angle information has a modest impact. Importantly, cross-generator tests reveal only minor performance degradation, suggesting that the trained models are largely generator-independent and potentially applicable to experimental data for rapid, model-stable event-shape estimation.

Abstract

This study demonstrates the application of supervised machine learning (ML) techniques to distinguish between isotropic and jet-like event topologies in heavy-ion collisions via the spherocity observable. State-of-the-art ML algorithms, optimized through systematic hyperparameter tuning, are employed to predict both traditional transverse spherocity $S_{0}$ and unweighted transverse spherocity $S_{0}^{p_{\rm T}=1}$ directly from raw event data. Moreover, the results from this study demonstrated that our approach remains largely model-independent, underscoring its potential applicability in future experimental heavy-ion physics analyses.

Figures (7)

• Figure 1: Illustration of weights utilization (represented by red arrows) during computation of $S_0$ (left) and $S_0^{p_{\rm T}=1}$ (right) in an jetty event.
• Figure 2: Correlation plots between the input features and the target variables. The values of Pearson correlation coefficient $r$ are also shown.
• Figure 3: ML models predictions versus true values of $S_0$ for 100 K testing data of minimum-bias Au+Au events at $\sqrt{s_{\rm NN}}=200$ GeV.
• Figure 4: ML models predictions versus true values of $S_0^{p_{\rm T}=1}$ for 100 K testing data of minimum-bias Au+Au events at $\sqrt{s_{\rm NN}}=200$ GeV.
• Figure 5: Comparison between the ML-predicted and true distributions of conventional spherocity (left) and unweighted spherocity (right) for minimum-bias Au+Au events at $\sqrt{s_{\rm NN}}=200$ GeV. The ratio of the model predicted values to the true values of spherocity are shown in the lower panels.