Addressing the Challenges of Heavy Ion Collision Parameters Estimation via Neural Network Techniques

Abstract We explore neural network techniques for determining fundamental parameters of heavy-ion collisions, focusing on two closely related objectives: the estimation of the impact parameter and the prediction of the number of spectator nucleons using information from a detector with small acceptance. Deep learning models trained with domain adaptation strategies and mixed datasets from multiple event generators (QGSM, EPOS, and PHQMD) demonstrate stable performance under limited and partially decorrelated detector information. We also extend our previously developed Deep Reconstruction Neural Network (DRNN) model to estimate the number of spectators. Results obtained for QGSM and EPOS indicate consistent behavior across these generators, while validation on PHQMD and full detector simulations are planned for future stages.