Multi-Modal Track Reconstruction using Graph Neural Networks at Belle II

This paper presents a multi-modal graph neural network that improves Belle II track reconstruction efficiency and purity by simultaneously processing data from both the central drift chamber and the silicon vertex tracker to mitigate the effects of high backgrounds and detector ageing.

The Gist

Imagine you are trying to solve a massive, high-speed jigsaw puzzle, but there’s a catch: the puzzle pieces are scattered across two different rooms, and the pieces themselves are constantly changing shape.

This is essentially the problem scientists at the Belle II experiment are facing. They are trying to track tiny, subatomic particles flying through a massive detector at incredible speeds. To do this, they use two main "cameras": the SVD (a high-precision silicon detector) and the CDC (a large gas-filled chamber).

Here is a breakdown of the paper’s breakthrough using a simple analogy.

The Old Way: The "Two-Room" Problem

Previously, the scientists used a "staged" approach. Imagine you have two detectives:

• Detective SVD stays in the small, high-tech room and tries to piece together clues.
• Detective CDC stays in the large, cavernous room and does the same.

The problem? Once they both finish their work, they have to meet in the hallway and try to "match" their findings. They’d say, "I found a clue in my room; does it belong to your clue?"

Because the rooms are so different (one uses light/silicon, the other uses gas/timing), they often make mistakes. They might accidentally link a clue from Room A to a completely different clue from Room B, creating a "fake" story (low purity), or they might fail to realize two clues actually belong to the same person (low efficiency). As the particles move further away from the center, this "matching" process becomes a total mess.

The New Way: The "Super-Brain" (BAT Finder)

The researchers created something called the BAT Finder (Belle II AI Track Finder). Instead of two separate detectives trying to talk to each other later, they built a single "Super-Brain" using a Graph Neural Network (GNN).

Think of the BAT Finder as a single, all-seeing eye that looks at both rooms at the exact same time.

Instead of treating the detectors as separate stages, the AI treats every single "hit" (a tiny piece of evidence) as a point on a giant, interconnected web. It doesn't care which room the clue came from; it just looks at the patterns in the web to see which points naturally "clump" together to form a path.

How the "Super-Brain" Works (The Tech)

The paper uses two clever tricks to make this brain work:

1. Multi-Modal Learning: Just like you can identify a song by both its melody and its lyrics, the AI looks at different types of data (timing, charge, position) from both detectors simultaneously.
2. Object Condensation: Imagine throwing a handful of glitter into the air. Instead of trying to track every single speck, the AI looks for where the glitter is "clumping" together. These clumps represent the actual paths of the particles.

The Results: A Massive Upgrade

The results were like upgrading from a blurry old TV to a 4K cinema screen:

• Better Catching (Efficiency): The old system was only catching about 48% of the particles that were moving in weird, displaced paths. The new BAT Finder caught 74.7%. It’s much better at finding the "hidden" particles.
• Fewer Mistakes (Purity): The old system often hallucinated "fake" particles by accidentally mixing up clues. The new system is incredibly clean, with a 97.6% accuracy rate in making sure the tracks it finds are real.

Why does this matter?

As these particle accelerators get more powerful, they create more "noise" (background interference). If we can't tell the difference between a real particle and random noise, we might miss the discovery of a lifetime—like a new particle that explains how the universe works. The BAT Finder gives scientists a much clearer lens to see through the chaos.

Technical Summary

Technical Summary: Multi-Modal Track Reconstruction using Graph Neural Networks at Belle II

1. Problem Statement

The Belle II experiment at the SuperKEKB collider faces significant challenges in charged-particle tracking due to high beam-induced backgrounds and detector aging. The current tracking system relies on a heterogeneous, multi-stage reconstruction chain involving the Silicon Vertex Detector (SVD) and the Central Drift Chamber (CDC).

The existing "Baseline" approach is sequential: it performs CDC track finding, extrapolates hits inward to the SVD using a Combinatorial Kalman Filter (CKF), performs SVD track finding, and then extrapolates back to the CDC. This staged process is difficult to optimize and prone to mismatches and double reconstructions, which degrade both track purity (the fraction of real tracks found) and track efficiency (the ability to find all particles), particularly for displaced tracks (particles originating far from the interaction point).

2. Methodology: The BAT Finder

To address these inefficiencies, the authors propose the Belle II AI (BAT) Finder, a unified, multi-modal Graph Neural Network (GNN) architecture.

• Multi-Modal GNN Architecture: Unlike previous staged models, the BAT Finder treats all hits from both the SVD and CDC as nodes in a single, unified graph. This allows the network to learn relationships between heterogeneous detector geometries without explicit subdetector ordering.
• Input Features:
  • SVD hits: Global coordinates, deposited charge, timing, and signal-to-noise ratios.
  • CDC hits: Drift time, time-over-threshold, signal amplitude, and wire-geometry information.
  • Handling Heterogeneity: Features are mapped into a common latent space using fully connected layers. Missing features for a specific detector are zero-padded, and a binary flag identifies the detector origin.
• Object Condensation: The model employs an "object condensation" loss function. Instead of traditional hit-linking, the network predicts:
  1. Cluster-space coordinates: To group hits belonging to the same particle.
  2. Condensation score ($\beta$): To identify valid track candidates and weight the importance of hits.
  3. Track parameters: Initial momentum, starting position, and charge.
• Training & Inference: The model is trained on realistic simulations including beam-induced noise. During inference, hits with high $\beta$ scores are selected as seeds, and hits are assigned to tracks based on their proximity in the learned cluster space. The resulting candidates are then refined using the standard GENFIT2 Kalman filter.

3. Key Contributions

• Unified Reconstruction: Replaces the complex, multi-stage SVD-CDC matching logic with a single-step inference process.
• Heterogeneous Data Integration: Successfully combines 3D silicon strip measurements with 2D drift-time measurements within a single GNN framework.
• Robustness to Displacement: Specifically designed to handle tracks with varying transverse displacements (up to 100 cm), which are traditionally difficult for sequential algorithms.
• Improved Feature Utilization: Incorporates CDC time-over-threshold information to better distinguish signal hits from background noise.

4. Results

The performance was evaluated using simulated displaced muon tracks. The BAT Finder significantly outperformed both the Baseline Finder and the previous CAT Finder (which was CDC-only):

Metric	Baseline Finder	CAT Finder	BAT Finder (Proposed)
Track Efficiency	48.0%	68.5%	74.7%
Track Purity	92.2%	93.8%	97.6%

• Efficiency Gain: The BAT Finder showed a consistent efficiency advantage across the entire displacement range. It particularly excels at small and intermediate displacements where the CAT Finder struggles due to its reliance on staged matching.
• Purity Improvement: The jump to 97.6% purity demonstrates that the unified approach effectively eliminates the "fake tracks" and duplicate reconstructions caused by failed detector-to-detector matching in the baseline method.

5. Significance

The BAT Finder represents a paradigm shift in Belle II tracking from sequential, heuristic-based matching to a holistic, machine-learning-driven approach. By resolving the inefficiencies of subdetector matching, the algorithm provides a robust solution for maintaining high-performance tracking in the face of increasing backgrounds. This work offers a scalable path for future high-luminosity operations at SuperKEKB, ensuring precise vertexing and momentum determination for physics searches beyond the Standard Model.