TL;DR
The Belle II experiment faces a significant challenge from increased background noise due to higher luminosity. This paper proposes a novel neural network-based Level 1 (L1) trigger system to accurately determine the z-vertex position of particle interactions in real-time. This approach promises to effectively suppress background events originating from outside the interaction region, a critical requirement for the experiment's success.
What Problem Does This Paper Address?
The SuperKEKB collider, which will host the Belle II experiment, is designed for unprecedented luminosity. This high luminosity, while enabling a vast number of physics events, also dramatically increases background noise. A primary source of this background is Touschek scattering, which generates events with vertices displaced from the intended interaction point. The predecessor, Belle, lacked a fast z-vertex trigger, leading to a poor signal-to-noise ratio for such events. Belle II requires a sophisticated L1 trigger system capable of rejecting these background events within a strict latency of 5 microseconds. Specifically, this paper tackles the challenge of precisely determining the longitudinal (z) position of event vertices in real-time to filter out these problematic background tracks.
Methods
The core of the proposed solution is a neural network (NN) based z-vertex trigger. Instead of reconstructing full tracks, the system leverages information from the Central Drift Chamber (CDC) within specific "sectors" defined by angular and momentum ranges. A specialized Multi-Layer Perceptron (MLP) is trained to estimate the z-vertex position directly from drift times within these sectors.
The approach involves a two-step prediction chain:
- Preprocessing: This stage uses information from the standard 2D trigger (providing track momentum and angle estimates) and a "preprocessing" MLP to estimate the polar angle ($\theta$) and a preliminary z-vertex position. This step helps to narrow down the search space.
- Expert MLPs: Based on the output of the preprocessing step, an appropriate "expert" MLP is selected. These expert MLPs are trained on specific, smaller sectors of the phase space (defined by $\phi$, $\theta$, and $p_T$) and provide a highly accurate z-vertex prediction.
The training of these MLPs utilizes an improved backpropagation algorithm called iRPROP, which is known for faster convergence. The implementation is designed for programmable hardware (FPGAs) to exploit the inherent parallelism of neural computation and meet the stringent latency requirements.
Datasets
The paper describes training and testing MLPs using simulated Monte Carlo (MC) data for the Belle II detector. These simulations incorporate realistic physics effects, including interactions with detector materials and non-linear drift time relationships. The authors specifically mention using MC data to define "relevant" Track Segments (TS) for sectorization and to evaluate the performance of their prediction chain. No specific benchmark datasets are introduced; rather, the focus is on simulating the Belle II environment.
Key Insights and Ideas
- NNs as a viable L1 trigger component: The paper demonstrates that neural networks, despite their computational complexity, can be effectively implemented in hardware for real-time trigger decisions, offering a robust alternative to traditional algorithms.
- Sectorization for efficiency: Dividing the complex phase space into smaller, specialized "sectors" and training dedicated "expert" MLPs for each sector is crucial for achieving high accuracy and managing computational load.
- Two-step prediction chain: The proposed preprocessing followed by expert MLPs provides a hierarchical approach that effectively reduces the search space and improves overall z-vertex resolution.
- Hardware implementation feasibility: The use of FPGAs (specifically the Xilinx Virtex-7) is highlighted as a promising solution for the parallel processing demands of the NN-based trigger.
Key Results
The study demonstrates that the proposed NN-based z-vertex trigger can achieve the required resolution.
- Z-vertex resolution: For single simulated muon tracks in narrow sectors, resolutions of RMS = 1.72 cm (low $p_T$) and RMS = 1.44 cm (high $p_T$) were achieved, well below the anticipated 2 cm requirement.
- Prediction chain performance: A three-step prediction chain achieved a final z-resolution of 1.7 cm.
- Efficiency: The ROC curves show significant improvements with each step of the prediction chain, reaching an efficiency of 98.2% with a contamination of 17.1% for a specific cut.
- Latency: Preliminary simulations suggest that the NN processing time, including parameter retrieval, can be kept under 400 ns, well within the L1 trigger budget.
Limitations and Future Work
- Background contamination: While the study demonstrates good resolution for signal-like tracks, the authors acknowledge that the real background rejection in the experiment will depend on the topological distribution of actual background tracks, which is still under investigation.
- Full event testing: The current studies are primarily based on single simulated tracks. Future work will involve testing the trigger system with full simulated events, including the combination of multiple track vertices.
- Algorithm optimization: Further research is planned to explore optimal algorithms for sector finding and the combination of different predictors in the preprocessing steps.
- Hardware implementation: The authors aim to install the neural network trigger into the Belle II trigger system and will continue to explore hardware solutions.
- Generalizability: The proposed method is noted to be applicable to other track-based trigger algorithms in experiments like the LHC and for future silicon detectors.
Summary updated