BSc thesis topics
FAQ - Informationen zur Bachelorarbeit in der Arbeitsgruppe von Prof. Ferber
| Topic: | Neural Network Compression and Quantization for the Belle II Level-1 Hardware Trigger |
| Summary: | Modern particle detectors such as the Belle II experiment rely on real‑time data processing to identify interesting collisions within microseconds through the Level‑1 hardware trigger. With increasing luminosity and background rates, efficient background filtering has become essential for maintaining the quality of trigger decisions. This motivates the use of compact and fast neural network models implemented on FPGAs for online hit filtering. This project investigates methods to reduce the computational complexity of such models through neural network quantization and compression, focusing on graph neural networks applied to hit filtering in the Central Drift Chamber of Belle II. The work involves exploring and comparing state‑of‑the‑art quantization‑aware training techniques (e.g. Brevitas, PQuant, NeuraLUT, HGQ, FitCompress) to achieve an optimal balance between model accuracy, inference latency, and hardware resource consumption. |
| You learn: | machine learning model optimization, quantization-aware training, model compression and hardware-aware design workflows |
| Prerequisites: | Good proficiency in English is required. Solid knowledge of Python programming, for example from the lecture and all exercises in Rechnernutzung, is required. The exam in ModEx 1+2 must be completed. Prior experience in software development projects is an advantage. Basic familiarity with machine learning frameworks such as TensorFlow or PyTorch is beneficial. |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Greta Heine |
| Last update: | 09.01.2026 |
| Topic: | Graph Neural Network Model Simplification and Regularization Strategies |
| Summary: | Real-time data processing in high-energy physics experiments, such as Belle II, requires machine learning models that are both accurate and computationally efficient. To meet the stringent timing and resource constraints of hardware trigger systems, models must often be simplified without sacrificing performance. This project investigates structural model reduction methods including pruning, low‑rank factorization, transferred or compact convolutional filters, and various forms of regularization such as L1 penalties and knowledge distillation. The goal is to develop lightweight neural networks that maintain high predictive power while meeting the latency and resource limits of FPGA‑based trigger hardware. |
| You learn: | machine learning model optimization, implementation and evaluation of pruning strategies and regularization methods, quantitative analysis of model complexity and inference performance, and hardware-aware design workflows |
| Prerequisites: | Good proficiency in English is required. Solid knowledge of Python programming, for example from the lecture and all exercises in Rechnernutzung, is required. The exam in ModEx 1+2 must be completed. Prior experience in software development projects is an advantage. Basic familiarity with machine learning frameworks such as TensorFlow or PyTorch is beneficial. |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Greta Heine |
| Last update: | 09.01.2026 |
| Topic: | GNN-Based ROI Finding in the Pixel Detector Using SVD Track Extrapolation at Belle II |
| Summary: | Track reconstruction at Belle II begins in the outer detectors, where occupancy is lower and pattern recognition is more efficient. An existing one-shot GNN-based algorithm for track finding in the Silicon Vertex Detector (SVD) provides a first reconstruction of charged particle trajectories. This information can be used to define a Region of Interest (ROI) in the innermost Pixel Detector (PXD), significantly reducing the combinatorial complexity in subsequent reconstruction steps. This project focuses on using SVD tracks reconstructed by the GNN algorithm to define ROIs in the PXD. You will develop and evaluate methods for projecting GNN-reconstructed tracks into the PXD, identifying the spatial volume in which associated hits are expected. The goal is to efficiently isolate relevant PXD hits, suppress background, and provide a clean input for downstream track fitting and vertex reconstruction. Particular emphasis will be placed on performance under high background conditions and integration into the existing Belle II software framework. |
| You learn: | GNN-based track reconstruction, detector geometry and projection techniques, ROI definition under realistic detector conditions, integration of machine learning with classical reconstruction algorithms |
| Prerequisites: | Good proficiency in English is required. Solid knowledge of Python programming, for example from the lecture and all exercises in Rechnernutzung, is required. The exam in ModEx 1+2 must be completed. Prior experience in software development projects is an advantage. Basic familiarity with machine learning frameworks such as TensorFlow or PyTorch is beneficial. |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Tristan Brandes |
| Last update: | 07.08.2025 |
| Topic: | GNN-Based Track Finding on Remaining Hits After Conventional Reconstruction at Belle II |
| Summary: | Conventional tracking algorithms at Belle II are highly optimised for efficiency but may miss tracks in high-occupancy or low-momentum regions, or in events with complex topologies. These missed tracks can have a significant impact on physics analyses, particularly those involving secondary vertices or long-lived particles. This project investigates a hybrid approach in which graph neural network (GNN) based track finding is applied after the standard reconstruction, using the remaining, unassigned hits. You will integrate a GNN tracking pipeline into the Belle II reconstruction framework to operate as a second-stage algorithm. The focus will be on identifying tracks that evade conventional pattern recognition, improving overall reconstruction efficiency. Tasks include defining suitable input data from remaining hits, training and optimising the GNN for this use case, and evaluating performance with respect to track recovery rates and reconstruction quality. |
| You learn: | Hybrid tracking strategies, advanced GNN architectures for particle physics, integration of machine learning with conventional algorithms, reconstruction efficiency optimisation |
| Prerequisites: | Good proficiency in English is required. Solid knowledge of Python programming, for example from the lecture and all exercises in Rechnernutzung, is required. The exam in ModEx 1+2 must be completed. Prior experience in software development projects is an advantage. Basic familiarity with machine learning frameworks such as TensorFlow or PyTorch is beneficial. |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Lea Reuter |
| Last update: | 07.08.2025 |
| Topic: | Slow Control Software for the AI-Based Calorimeter Trigger CaloClusterNet in Belle II |
| Summary: | The AI-based calorimeter trigger at Belle II requires precise control over hardware parameters to ensure stable and efficient operation. A robust slow control system is essential for configuring trigger settings, monitoring system health, and responding to environmental changes in real time. In this project, you will develop slow control software in C/C++ to manage and optimize the AI-based calorimeter trigger. This includes implementing configuration interfaces, integrating hardware monitoring tools, and designing automated response mechanisms to maintain stable trigger performance. Your work will directly contribute to the reliable operation of this state-of-the-art trigger system in a high-energy physics experiment. |
| You learn: | FPGA slow control |
| Prerequisites: | good proficiency in English is required, good knowledge in C/C++ is required, exam in ModEx1+2 is a plus, first experience in own software projects is a plus |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Marc Neu |
| Last update: | 04.06.2025 |
| Topic: | GNN-Based Track Reconstruction in the Silicon Pixel Detector of Belle II |
| Summary: | The Belle II Pixel Detector (PXD), located closest to the interaction point, presents unique challenges due to its small pixels and extreme background conditions. Efficient track reconstruction in this environment is crucial for precise vertex determination in Belle II. This project will extend the CATFinder ) to incorporate information from the high-resolution PXD. You will develop a GNN-based approach that integrates PXD data into the existing tracking framework, optimizing it for the high-occupancy conditions near the beam pipe. Your work will focus on improving robustness against background hits, refining pattern recognition, and enhancing overall tracking efficiency in the inner detectors of Belle II. |
| You learn: | advanced track reconstruction techniques in particle physics, machine learning |
| Prerequisites: | good proficiency in English is required, good knowledge in Python programming (e.g. lecture and all exercises in "Rechnernutzung") is required, exam in ModEx1+2 is required, first experience in own software projects is a plus, basic knowledge in machine learning tools (e.g. Tensorflow or PyTorch) is a plus |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Lea Reuter |
| Last update: | 04.06.2025 |
| Topic: | Automated process monitoring with workflow management software towards sustainable computing |
| Summary: | In High Energy Physics many calculations are done on complex data. Since every calculation increases energy consumption, it is important to do these as sensible and efficient as possible. While some computing sites already provide data on the power usage of these calculations, it is still difficult to monitor the energy consumption of whole analysis workflows. In this project, an existing workflow management based on the software luigi will be extended to automatically collect data on the energy usage of its tasks. |
| You learn: | workflow management software (luigi), monitoring and accounting of computing resources, databases |
| Prerequisites: | good proficiency in English is required, good knowledge in Linux is required, exam in ModEx1+2 is required, first experience in own software projects is a plus |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Jonas Eppelt |
| Last update: | 04.06.2025 |
| Topic: | Performance optimization for Machine Learning reconstruction algorithms |
| Summary: | This thesis project centers on optimizing the performance of Machine Learning (ML) reconstruction algorithms, particularly focusing on Python algorithms and their integration with C++ interfaces for the high-level trigger at Belle I running on a large computing cluster. As a bachelor student, your primary objective will be to streamline the execution of ML reconstruction algorithms on CPUs and GPUs. You will explore techniques such as algorithm parallelization, memory management, and code optimization to achieve optimal performance in both Python and C++ environments. Through rigorous benchmarking and profiling, you will evaluate the impact of your optimizations on the reconstruction speed and resource utilization. By the end of your thesis, you will have contributed to the development of robust and efficient ML reconstruction pipelines, essential for high-level trigger systems in particle physics experiments. |
| You learn: | advanced track reconstruction techniques in particle physics, advanced C++ optimization |
| Prerequisites: | good proficiency in English is required, good knowledge in Python and C++ programming (e.g. lecture and all exercises in "Rechnernutzung") is required, exam in ModEx1+2 is required, first experience in own software projects is a plus, basic knowledge in machine learning tools (e.g. Tensorflow or PyTorch) is a plus |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Dr. Giacomo De Pietro |
| Last update: | 04.06.2025 |
| Topic: | Performance study of NN-Training with data streaming from remote storage |
| Summary: | In High Energy Physics, datasets of several hundreds of GB are used for machine learning. Often, these datasets are stored at remote storage servers. Sometimes it is not possible to keep them on the computing machine. It could be an option to stream the data from the remote storage. It is possible to use a cache that keeps the data near the computing resource. That reduces the network traffic and the access latency for the files. The performance during streaming with and without cache should be studied. |
| You learn: | concepts and usage of distributed computing, transfer protocol, and storage/caching software XRootD |
| Prerequisites: | good proficiency in English is required, good knowledge in Linux is required, exam in ModEx1+2 is required, first experience in own software projects is a plus, basic knowledge in machine learning tools (e.g. Tensorflow or PyTorch) is a plus |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Dr. Matthias Schnepf |
| Last update: | 04.06.2025 |
| Topic: | Multi-core support for Belle II Grid |
| Summary: | Belle II uses a distributed computing infrastructure called Grid. Currently, Belle II supports only one CPU core per application in the Grid. The use of multiple CPU cores would speedup applications and reduce memory footprint. You can improve the Belle II Grid computing by configuring, implementing, and testing the support for multiple CPU cores in the Belle II Grid. |
| You learn: | usage and configuration of the workflow management system DIRAC, Python, bash, Grid computing |
| Prerequisites: | good proficiency in English is required, good knowledge in Linux is required, exam in ModEx1+2 is required, first experience in own software projects is a plus |
| Supervisor: | Prof. Dr. Torben Ferber |
| Contact: | Dr. Matthias Schnepf |
| Last update: | 04.06.2025 |