以卷積神經網路分離 Belle II 電磁量能器中重疊光子射叢

Abstract

在 Belle II 電磁熱量計中，當兩個粒子射叢十分靠近，能量可能會重疊，而能量較難分配給兩個粒子射叢。在蒙特卡羅模擬下，粒子射叢被劃分為5×5扣掉4 個角的CsI晶體。輸入能量圖並使用卷積神經網絡（CNN），選擇2GeV以下的射叢能量分佈，當作數據集作為訓練模型，而4GeV的能量作為測試已建立模型的數據集。主要測試不同結構的CNN模型對能量解析度的影響。此外，將全連接網絡作為一個簡單的模型進行測試，並將 LeNet 和 AlexNet 作為一種著名的 CNN 進行測試，目的是提高能量解析度。

In the Belle II electromagnetic calorimeter (ECL), two particle showers close together, the energy may be overlap. It is difficult to separate the energy. Under Monte Carlo simulation, the shower is grouped into 5×5 CsI crystals without 4 corners. By the image map of the energies, convolutional neural network (CNN) is used to split the photon shower. Choose the energy distribution of the showers under 2 GeV, the dataset is concerned as the training model. The energy of 4 GeV as the dataset that test the model. Mainly test the influence of CNN models of different structures on energy resolution. In addition, test Fully-connected network as a simple model, and test LeNet and AlexNet as a famous kind of CNN. The purpose is to improve the energy resolution.