Determination of impact parameter in high-energy heavy-ion collisions via deep learning

TL;DR

This paper demonstrates that deep learning models, specifically DNN and CNN, can accurately predict impact parameters in high-energy heavy-ion collisions from final-state particle spectra, with CNN showing superior performance especially at larger pseudorapidity ranges.

Contribution

It introduces the application of deep neural networks to determine impact parameters from collision data, highlighting CNN's improved accuracy and interpretability via Grad-CAM.

Findings

Both DNN and CNN achieve about 0.4 fm mean absolute error.

CNN performs better at larger pseudorapidity ranges.

Models are effective across different beam energies.

Abstract

In this study, Au+Au collisions with the impact parameter of $0 \leq b \leq 12.5$ fm at $\sqrt{s_{NN}} = 200$ GeV are simulated by the AMPT model to provide the preliminary final-state information. After transforming these information into appropriate input data (the energy spectra of final-state charged hadrons), we construct a deep neural network (DNN) and a convolutional neural network (CNN) to connect final-state observables with impact parameters. The results show that both the DNN and CNN can reconstruct the impact parameters with a mean absolute error about $0.4$ fm with CNN behaving slightly better. Then, we test the neural networks for different beam energies and pseudorapidity ranges in this task. It turns out that these two models work well for both low and high energies. But when making test for a larger pseudorapidity window, we observe that the CNN shows higher prediction accuracy than the DNN. With the method of Grad-CAM, we shed light on the `attention' mechanism of the CNN model.