Neural networks for impact parameter determination *
Hartnacktg,Basstf,Maruhni,H. Stiickert,W. Greinerf,Reinhardtt,Bischofft
Abstract:A c c m e impacl parameter determination in a heavy-ion collision is crucial for almost all further analysis. We investigate the capabilities of an attificial neural network i n lhst respeck First resulls show lhal the neural network is capable of improving the accuracy of the impact parameter determination based on observables such as the Row angle, the average directed inplane uansvene momentum and the difference between lnnwerse and longitudinal momenla. However. further investigations are necessary U) discover the full potential of lhe mural network approach. The physics of heavy-ion collisions is motivated by the idea to leam something about the properties of hot and dense nuclear matter [MI. In order to investigate highly compressed nuclear matter in heavy-ion collisions it is important to select only the most central collisions. On the other hand, recently discovered new phenomena such as pionic bounce-off and squeeze-out are only observed in semiperipheral collisions. Therefore the proper determination of the impact parameter in a heavy-ion collision is crucial to almost all fuither analysis. There have been various approaches towards the determination of the impact parameter, most of which are either based on the mean particle multiplicity or on a transverse momentum analysis (directivity-cut). However, all of these methods have one thing in common: they tend to break down for impact parameters smaller than 2 fenni. In general the accuracy is about dz1 to f1.5 fermi. In this contribution, an artificial neural network has been used to determine the impact parameter. For a proper analysis of the network's performance, the analysed heavy-ion collisions have to be supplied by a theoretical event-generator rather than by experiment. Otherwise it would be impossible to compare the network output with a target value for the impact parameter of the heavy-ion collision. For our investigations, we applied an extension of the quantum molecular dynamics model (QMD) [9-l1], which explicitly incorporates isospin and pion production via the delta resonance (IQMD) [ 12-14]. In the QMD model the nucleons are represented by Gaussian-shaped density distributions. They are initialized in a sphere of a radius R = I .14A'I3 fm, according to the liquid-drop model. Each nucleon is supposed to occupy a volume of h3, so that the phase space is uniformly filled. The initial momenta are randomly chosen between 0 and the local Thomas-Fermi momentum. The A p and AT nucleons interact via twoand three-body Skyrme forces, a Yukawa potentid. ' Supponed by GSI. BMFT and DFG 0954-38YY/94fl10021+06$0750 Q 1994 IOP Publishing Lld L21 L22 Lefter to the Edifor momentum-dependent interactions, a symmetry potential (to achieve a correct distribution of protons and neutrons in the nucleus) and explicit Coulomb foxes between the Zp and ZT protons. They are propagated according to Hamilton’s equations of motion. Hard N-N collisions are included by employing the collision term of the well known vuu/Buu equation [ 5 , 15-18]. The collisions are done stochastically, in a similar way as in the cascade models [ 19,201. In addition, the Pauli blocking (for the final state) is taken into account by regarding the phase-space densities in the final states of a two-body collision. As input for the neural net we use the flow angle Sflow, the average directed inplane transverse momentum Pr,dit, and an observable Q Z Z which compares transverse with longitudinal momentum. The flow angle can be extracted from the flow-tensor
Physics,Computer Science