Estimating centrality in heavy-ion collisions using Transfer Learning technique

TL;DR

This paper demonstrates that transfer learning with pre-trained CNN models can effectively estimate collision centrality in heavy-ion collisions, achieving high accuracy even with models trained on unrelated datasets.

Contribution

The study applies transfer learning with popular CNN architectures to estimate collision centrality in heavy-ion collisions, showing promising results and highlighting the potential of this approach.

Findings

VGG16 outperforms other models in estimating $N_{\rm part}$

All models achieved good performance despite being pre-trained on different domains

Transfer learning can extract meaningful observables from heavy-ion collision data

Abstract

In this study, we explore the applicability of Transfer Learning techniques for estimating collision centrality in terms of the number of participants ($N_{\rm part}$) in high-energy heavy-ion collisions. In the present work, we leverage popular pre-trained CNN models such as VGG16, ResNet50, and DenseNet121 to determine $N_{\rm part}$ in Au+Au collisions at $\sqrt{s}=200$ GeV on an event-by-event basis. Remarkably, all three models achieved good performance despite the pre-trained models being trained for databases of other domains. Particularly noteworthy is the superior performance of the VGG16 model, showcasing the potential of transfer learning techniques for extracting diverse observables from heavy-ion collision data.