Background event classification for Belle II

In this project I used neural networks to classify EvtGen simulated data into signal, \(B \to K^* \nu \bar{\nu} \), and background. The goal was to investigate the possibility of bypassing the expensive Geant4 detector simulation and following analysis. This project was done in the realms of the Belle-II experiment, but is applicable to all of experimental particle physics. It was a very new direction in the collaboration, where I helped to develop one of the first convolutional neutral networks for classification. By restructuring the network layers and managing to constructively input additional data (such as the decay string), I managed to improve the accuracy by over 10% and set the groundwork for future contributions. The project is still ongoing and being constantly improved.

During this project I learned about neural networks, some common design practises and options for handling and pre-processing data. I used Keras and other parts of Tensorflow to embed relevant data from the simulation and develop a convolutional neural network for classification of signal and background in particle collisions. Furthermore, I learned about the simulation side of particle physics and some details about general data handling.