Masterarbeiten zu Software
| Thema: | Real-time Graph Neural Networks for the Belle II calorimeter |
| Zusammenfassung: | The GNN-ETM* team at KIT ETP and ITIV successfully operated an ultrafast clustering algorithm based on Graph Neural Networks (GNN) deployed on AMD Xilinx UltraScale hardware for the first time at the end of 2024 at the Belle II calorimeter (ECL). However, to integrate this solution into the real-time trigger system of the Belle II experiment by autumn 2025, the execution speed must be doubled, reaching a milestone of 1.5 microseconds. Positioned at the intersection of particle physics and electrical engineering, the candidate will collaborate with an interdisciplinary team at KIT ETP and ITIV, where the algorithms and hardware implementations are being developed. The first focus of the research is to optimize the current GNN architecture, particularly by exploring alternatives to the existing latent-space graph building GravNet layers. One approach is to replace these with network layers that eliminate the time-consuming sorting step while maintaining dynamic graph construction. Another possibility is to pre-build static graphs based on the real-space relationships between calorimeter Trigger Cells (TCs), defining edges according to the presence of TCs in an event. The feasibility of these approaches for hardware implementation will be evaluated in collaboration with electrical engineering students at ITIV, enabling an iterative process to co-design the optimal software/hardware solution. The second area of research focuses on improving the quantization process for GNN clustering. This involves exploring quantization-aware training techniques, including experimenting with frameworks like QONNX and refining quantization methods in QKeras. Key areas for improvement include adjusting quantization scales and types, and evaluating the impact of modifying activation functions, particularly at the output layers. Additionally, the thesis will investigate hardware-specific quantization methods to better align the model's behavior with the actual hardware constraints. *GNN-ETM: Graph Neural Network ECL Trigger Module |
| Sie lernen kennen: | Python programming, machine learning, trigger, algorithm optimization |
| Referent: | Prof. Dr. Torben Ferber |
| Ansprechpartner: | Isabel Haide (ETP) |
| Letzte Änderung: | 26.01.2025 |
| Thema: | Next generation AI for the upgrade of the Belle II calorimeter trigger |
| Zusammenfassung: | The Belle II experiment is preparing for a major upgrade to its calorimeter (ECL) readout and trigger system, which will involve utilizing a significantly higher granularity of data. This upgrade will also leverage the next generation of hardware, specifically the AMD Xilinx Versal AI Edge Series Gen 2 platforms, to meet the increased computational demands. The candidate will join the interdisciplinary GNN-ETM team at KIT ETP and ITIV, where both the algorithms and hardware implementations are being developed. The input data features high sparsity, meaning that most inputs will not provide significant information for the algorithm. The research will focus on adapting machine learning models, specifically Graph Neural Networks (GNNs), to work efficiently on AMD Xilinx Versal devices, including its AI-cores and powerful FPGA capabilities. The candidate will directly contribute to the upgrade project by investigating the required energy and timing resolution per input and by leveraging waveform information from individual calorimeter crystals. Current plans involve deploying multiple parallel Versal platforms to segment the calorimeter and parallelize the reconstruction process. The candidate will explore optimal segmentation patterns to ensure efficient data processing across platforms while meeting real-time constraints. Given that the AMD Xilinx Versal devices are cutting-edge real-time processors, close collaboration with electrical engineering students at ITIV will be essential to co-design the optimal software-hardware interface for the upgrade. The GNN-ETM team operates its own AMD Xilinx Versal devices at KIT, where the algorithms will be tested on real hardware to validate the proposed solutions. *GNN-ETM: Graph Neural Network ECL Trigger Module |
| Sie lernen kennen: | Python programming, machine learning, trigger, algorithm optimization |
| Referent: | Prof. Dr. Torben Ferber |
| Ansprechpartner: | Isabel Haide |
| Letzte Änderung: | 26.01.2025 |