CLUSTER-GRAVITATION - AN EXTENSION TO THE KOHONEN MAP FOR THE IDENTIFICATION OF pp-BREMSSTRAHLUNG AT COSY

A self-organizing network according to a modiied Kohonen algorithm was used for the identiication of 272 Events of pp-bremsstrahlung in a pp-scattering experiment at the SATURNE accelerator at T beam =1.0 GeV. 1. Identiication of the pp-bremsstrahlung One of the experimental goals being pursued by means of the TOF-spectrometer 1] at the COSY accelerator at the Forschungszentrum J ulich is the measurement of pp-bremsstrahlung (pp!pp) to get direct information on the nucleon-nucleon interaction. The TOF-spectrometer detects the two protons; the photon has to be reconstructed via four-momentum conservation. Background reactions which pass the trigger are pion production such as pp!pp 0 and nuclear reactions in the target foils, pA 1 !ppA 2. The standard ooline method for reaction identiication is a missing-mass analysis. The relative number of correctly identiied events of pp-bremsstrahlung by this method is 100% below the pion threshold (T beam =280 MeV), decreasing to 0% at T beam =600 MeV due to nite detector resolution. Since the missing-mass m 3 is a one-dimensional parameter, we tried neural networks to use kinematic information directly in a high-dimensional parameter space. However, deduced kinematic input parameters were E 3 , p 3 , # 3 , # 1 + # 2 and ' 12 (i=1,2 denoting the protons, i=3 the third particle). A feed-forward backpropagation network (FF network), which we formerly tried 2], was not applicable because of deviations of model-dependent Monte Carlo simulated training data from the real experimental data. In this paper we present a method for the identiication of pp-bremsstrahlung based upon a self-organizing network (SO network) according to the Kohonen algorithm 3] with its feature of unsupervized learning directly using model-independent experimental data. The SO network application was tested with data of a pp-and pA-scattering experiment (T beam =1 GeV) at SATURNE 4].