Summary of the Dissertation
Point cloud networks are well-suited for irregular detector geometries with millions of sensors, but to date there is no method for deploying them in real-time trigger systems. This paper introduces DeePloy, which maps graph-based point cloud networks to dataflow accelerators on FPGAs and heterogeneous SoCs under strict real-time constraints. In three practical applications within the Belle-II experiment, DeePloy demonstrates that such networks can meet the stringent real-time requirements of large-scale scientific experiments.
Data Flow Processing for DAQ Systems
New advances in telecommunications are leading to increasing bandwidth requirements in digital signal processing. New standards are capable of transmitting sub-terabit data rates. Characterizing such systems requires the processing, validation, and storage of the generated measurement data. Stream processors play a key role in the development of data acquisition systems (DAQ systems). At ITIV, we aim to develop a framework for reconfigurable DAQ systems that enables the specification of future transmission standards.
Implementation of GNNs on Hardware Accelerators
Graph Neural Networks (GNNs) extend conventional deep learning methods to graph structures. Their generalized formulation opens up new possibilities in application areas such as image processing, monitoring, and network analysis. Bandwidth and memory latency pose significant bottlenecks, particularly during the implementation of GNNs in real-time applications. Therefore, the use of reprogrammable hardware platforms such as FPGAs is a central topic in current research. At ITIV, we are working to improve the usability of GNNs in embedded systems.
Data-Driven Design of Trigger Systems
Particle accelerators generate enormous amounts of data during their experiments and therefore require so-called trigger systems. These systems implement filtering mechanisms to distinguish between relevant and irrelevant detector events. Varying hyperparameters during physics experiments require automated training and reconfiguration of the detector’s firmware. Here at ITIV, we are investigating measures for adapting latency-optimized trigger systems to changing environmental variables.
| Title | Type |
|---|---|
| Decoding Decay - Real-Time Particle Detection from Point Clouds | Masterarbeit |
| Compiler-Based Integration of Neural Network Accelerators | Masterarbeit |
| Title |
|---|
| Empowering Tomorrow's Engineers: An MLIR-Based Toolchain for Transforming Python Neural Networks into Verilog Hardware |
Publications
Neu, M.
2026, September 3. Karlsruher Institut für Technologie (KIT). doi:10.5445/IR/1000196764
Liu, Y.-X.; Koga, T.; Bae, H.; Yang, Y.; Kiesling, C.; Meggendorfer, F.; Unger, K.; Hiesl, S.; Forsthofer, T.; Ishikawa, A.; Ahn, Y.; Ferber, T.; Haide, I.; Heine, G.; Hsu, C.-L.; Little, A.; Nakazawa, H.; Neu, M.; Reuter, L.; Savinov, V.; Unno, Y.; Yuan, J.; Xu, Z.
2026. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1084, Art.-Nr.: 171248. doi:10.1016/j.nima.2025.171248
Neu, M.; Baptist, F.; Lobmaier, T.; Papagno, F.; Ferber, T.; Becker, J.
2026. 2026 IEEE 34th Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM), 289–293, Institute of Electrical and Electronics Engineers (IEEE). doi:10.1109/FCCM68464.2026.00074