Master Thesis Defense by Ian Pascal Møller

Title: Investigating the use and capabilities of neural networks for data analysis and anomaly detection in the ALICE hadronic forward calorimeter

Abstract: This thesis investigates the usage of convolutional neural networks to perform various analysis tasks on calorimetry data by considering the data as regular 2D images. The data is obtained from both experimental testing and simulations using the ALICE Forward Hadronic Calorimeter prototype, whose design means that we can easily and directly represent the data as 2D images, which we can then treat as a computer vision problem.

The purpose of the thesis is to explore the capabilities of neural networks within this field, and develop novel tools to aid in specific data analysis tasks that can outperform conventional algorithms. These tasks are divided into two groups: one which concerns counting, locating and clustering particle showers in events, and one which concerns the detection of anomalous data present in events. By using neural networks for these purposes, we aim to be able to address some of the shortcomings of the conventional methods and thus provide the foundations for better future analysis tools.

In the presentation, it will be shown that not only are these tasks all possible to accomplish, but also that a properly trained neural network can outperform conventional clustering algorithms significantly when it comes to counting and locating particle showers. Furthermore, we will show that autoencoder networks can be readily used to detect anomalies in data even when these anomalies are very small, which can greatly help troubleshooting potential issues both during and after data acquisition, as will be demonstrated through a practical case study.