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.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: how to use transfer learning techniques to estimate the centrality in high - energy heavy - ion collisions, especially with the number of participating nucleons ($N_{\text{part}}$) as an indicator. Specifically, the researchers hope to use transfer learning methods and pre - trained convolutional neural networks (such as VGG16, ResNet50 and DenseNet121) to determine the $N_{\text{part}}$ of each event in Au + Au collisions at $\sqrt{s}=200 \, \text{GeV}$.
### Problem Background
In high - energy heavy - ion collision experiments, the collision centrality is a key parameter used to quantify the geometry of the collision and the degree of nucleon overlap. Theoretically, the collision centrality can be described by parameters such as the impact parameter ($b$), the number of binary nucleon - nucleon collisions ($N_{\text{coll}}$) and the number of participating nucleons ($N_{\text{part}}$). However, these parameters cannot be directly measured in experiments and are usually derived from other measurable physical quantities (such as the charged - particle multiplicity or the forward energy carried by spectator nucleons).
### Research Motivation
Although traditional methods based on the Glauber model can estimate the collision centrality, they have limitations in accuracy and generalization ability. In recent years, machine - learning and deep - learning techniques have made significant progress in the field of high - energy physics, especially in estimating collision parameters. Therefore, this paper aims to explore the application potential of transfer learning techniques in this task.
### Main Contributions
1. **Application of Transfer Learning**: Use pre - trained CNN models (VGG16, ResNet50 and DenseNet121) for transfer learning to estimate $N_{\text{part}}$ in high - energy heavy - ion collisions.
2. **Model Performance Evaluation**: Evaluate the performance of different models through two methods, fixed - feature extraction and fine - tuning, and compare their performance in different collision centrality regions.
3. **Data Processing and Preparation**: Use the AMPT model to generate input data and convert the momentum - weighted ($\eta-\phi$) spectra of charged hadrons into 32×32 two - dimensional histograms as input features.
### Conclusion
The research shows that all three models (VGG16, ResNet50 and DenseNet121) show good performance in estimating $N_{\text{part}}$, especially the VGG16 model, which performs particularly well in terms of accuracy and robustness. The effectiveness of transfer learning techniques has been verified, providing new ideas and methods for further improving the estimation accuracy of collision centrality.