End-to-End Pipeline for Trigger Detection on Hit and Track Graphs

Tingting Xuan, Yimin Zhu, Giorgian Borca-Tasciuc, Ming Xiong Liu, Yu Sun, Cameron Dean, Yasser Corrales Morales, Zhaozhong Shi, Dantong YuAAAI 2023 · DOI

Introduction: The Challenge of Real-Time Data Filtering

In high-energy nuclear physics, experiments such as the sPHENIX detector at the Relativistic Heavy Ion Collider (RHIC) generate data at an immense scale. Operating like a 250-megapixel camera capturing 10 million collisions per second, the detector produces a volume of information that far exceeds storage capacities. The primary challenge lies in identifying rare, scientifically significant events—such as the decay of heavy charm quarks—amidst a overwhelming background of "uninteresting" collisions.

Traditionally, hardware and software "triggers" filter this data in real-time. However, deterministic triggers often struggle with the complexity of heavy-flavor decays. This research introduces an AI-enabled, end-to-end pipeline designed to automate these high-accuracy "keep/discard" decisions at line speed using Graph Neural Networks (GNNs).

Methods: A Multi-Stage GNN Pipeline

The proposed system processes raw detector readouts through a three-stage pipeline that transitions from low-level pixels to high-level physics decisions:

  1. Pixel Clustering: Raw pixel readouts are grouped into "hits" using a deterministic connected-component algorithm. By utilizing a hash table for neighbor discovery, the authors achieve $O(n)$ time complexity, a prerequisite for real-time operation.
  2. Track Reconstruction: The system connects hits across detector layers to reconstruct particle trajectories (tracks). A Message-Passing GNN performs link prediction to determine which hit pairs belong to the same particle trajectory.
  3. Trigger Detection: The core innovation is the transition from a "hit-graph" to a track-graph. In this representation, each node is a reconstructed track rather than a single coordinate. The system uses ParticleNet to dynamically generate an adjacency matrix for these tracks, identifying if they likely share a common origin (vertex). Finally, SAGPool (Self-Attention Graph Pooling) aggregates these features into a graph-level representation for binary classification.
The end-to-end pipeline from raw pixels to trigger detection.

To refine the model's physical accuracy, the authors introduced a weighted Laplacian Loss. This regularization term penalizes the model for predicting false connections between tracks that do not share a physical vertex, ensuring the learned graph structure aligns with the underlying geometry of particle decays.

Experimental Setup and Simulation

Due to the complexity of the sPHENIX environment, the pipeline was developed and validated using a large-scale simulation dataset. Particle collisions were generated using PYTHIA8.3, while GEANT4 was employed to simulate the interaction of those particles with the detector material.

The sPHENIX detector design and simulation environment.

The study specifically focused on the MVTX (Monolithic Active Pixel Sensor Vertex Detector) and INTT (Intermediate Tracker) subsystems. These components are critical for capturing the initial stages of particle decay and are the fastest subsystems within the sPHENIX architecture. The models were trained on 600,000 simulated events and validated on an additional 200,000.

Key Insights and Results

The shift from hit-graphs to track-graphs represents a significant conceptual leap. While hits are merely spatial coordinates, tracks represent physical entities. By focusing on inter-track relationships, the GNN mimics the logic human physicists use to identify interesting events. Furthermore, track-graphs are significantly sparser than hit-graphs, leading to faster inference and lower memory overhead.

Performance Metrics: * Tracking Accuracy: The GNN-based tracking stage correctly identified and connected hits into tracks with over 90% accuracy. * Trigger Performance: The track-graph approach achieved approximately 72% accuracy for trigger detection. * Comparative Advantage: The proposed method outperformed the baseline hit-graph approach by 2%. In a high-luminosity environment filtering 10 million events per second, a 2% improvement translates to a massive reduction in wasted storage and a higher purity of collected data.

Conclusion and Future Work

This end-to-end pipeline demonstrates the potential for GNNs to handle the extreme data rates of modern physics experiments. While the results in simulation are promising, real-world deployment will face challenges including detector noise and hardware latency. Future work will focus on providing detailed latency benchmarks and extending the architecture to detect a broader range of rare physical phenomena simultaneously as the system moves toward live production environments.

Summary updated

BibTeX
@inproceedings{xuan2023end,
  author = {Tingting Xuan and Yimin Zhu and Giorgian Borca-Tasciuc and Ming Xiong Liu and Yu Sun and Cameron Dean and Yasser Corrales Morales and Zhaozhong Shi and Dantong Yu},
  title = {{End-to-End Pipeline for Trigger Detection on Hit and Track Graphs}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.26870},
}
Figure