以卷積神經網路進行Belle II電磁量能器中重疊光子射叢之分離以及正負電子對撞實驗中雙粒子關聯性之測量

Abstract

本論文包含兩個主題。在第一個主題中我們利用機器學習的方式去改進Belle II軟體中重疊光子射叢分離的方法，而第二個主題則是使用Belle探測器中正負電子對撞的資料來做雙粒子關聯性的測量。 在Belle II的電磁量能器中，粒子射叢能用5*5大小的CsI晶體組來表示。而在某些情況下，複數個射叢的晶體組可能會相互重疊。我們預期經由研究晶體能量的分布，卷積神經網路的圖形辨識能力可以很好的進行重疊光子射叢的分離，因此我們以電磁量能器的能量分佈訓練卷積神經網路。經由大量的模擬資料進行訓練，我們預期神經網路模型可以改善重疊光子射叢角位置的解析度。 在第二個主題中，我們測量了Belle探測器中正負電子對撞產生的帶電粒子之雙粒子關聯性，其中正負電子的質心能量為10.58 以及 10.52 GeV。關聯性函數測量了大範圍贋快度與完整範圍方位角，並以帶電粒子多樣性為函數。藉由同時比較不同事件產生器的模擬結果，這份研究揭露了測量結果與模擬的差異。

In this thesis, two major topics are studied. The first topic is to correct the photon shower splitting algorithm in the Belle II software with machine learning method. The second topic is the study about two-particle correlation function performed with Belle e+e- collisions data. In the Belle II electromagnetic calorimeter (ECL), each particle shower can be described by a set of 5*5 CsI crystals. In some case, the crystal sets of several showers will be overlapped with each other. We expect that by studying the distribution of crystal energies, the photon shower splitting algorithm can be improved with the ability of pattern recognition in convolutional neural network (CNN). Therefore, we trained the CNN model with ECL crystal energies distribution as input image. With a large amount of simulation data, the model is expected to improve the resolution of shower angular position. In the second topic, we present the measurements of two-particle angular correlations for charged particles in e+e- collisions with the Belle detector at a center-of-mass energy of 10.58 and 10.52 GeV. The correlation functions are measured over a broad range of pseudorapidity and full range of azimuthal angle as a function of charged particle multiplicity. With comparing the results to the expectations of several event generators, it reveals the discrepancies between the measurements and predictions.