Noise waveform generation using GANs and charged particle identification using pulse shape discrimination in the belle II electromagnetic calorimeter

Abstract

This thesis investigates the use of generative adversarial networks (GANs) as an alternative method to simulate noise waveforms for Belle II CsI(Tl) calorimeter crystals. Presented is a deep convolutional GAN (DCGAN) trained using background waveforms recorded in the ECL during a physics run. Results are presented showing good agreement in the distribution of metrics comparing data and simulated noise waveforms using a two-sample Kolmogorov-Smirnov test. The models are shown to be difficult to train, and many possible improvements are identified. Secondly, this thesis showcases the development of a particle identification tool relying on pulse shape discrimination (PSD) as an input to a gradient boosted decision tree (GBDT) classifier. Two models are trained to discriminate μ±, π± and e±, π±. Results show that PSD charged particle identification in the ECL improves the e±, π± discrimination, but result in smaller improvements to the μ±, π± discrimination. Results also show an improvement to the result obtained with the currently implemented PSD discriminator trained on neutral particles (γ, K-long).