Estimation of collision centrality in terms of the number of participating nucleons in heavy-ion collisions using deep learning

Abstract

The deep learning technique has been applied for the first time to investigate the possibility of centrality determination in terms of the number of participants ($N_{\mathrm{part}}$) in high-energy heavy-ion collisions. For this purpose, supervised learning using both deep neural network (DNN) and convolutional neural network (CNN) is performed with labeled data obtained by modeling relativistic heavy-ion collisions utilizing A Multi-phase Transport Model (AMPT). Event-by-event distributions of pseudorapidity and azimuthal angle of charged hadrons weighted by their transverse momentum are used as input to train the DL models. The DL models did remarkably well in predicting $N_{\mathrm{part}}$ values with CNN slightly outperforming the DNN model. The Mean Squared Logarithmic Error (MSLE) for the CNN model (Model-4) is determined to be 0.0592 for minimum bias collisions and 0.0114 for 0-60\% centrality class, indicating that the model performs better for semi-central and central collisions. Furthermore, the studied DL model is proven to be robust to changes in energy as well as model parameters of the input. The current study demonstrates that the data-driven technique has a distinct potential for determining centrality in terms of the number of participants in high-energy heavy-ion collision experiments.