Mitigating Detector Ageing Effects with Graph-Based Multi-Modal Track Reconstruction at Belle II

Imagine a giant, high-speed camera inside a particle accelerator, designed to take pictures of the universe's tiniest building blocks. This camera, called a detector, is made of thousands of tiny wires that act like a net, catching the invisible paths of particles as they zoom by. For this camera to work perfectly, it needs to be clean, the wires need to be strong, and the "film" (the data) needs to be complete. But here's the problem: the machine is so powerful that it's slowly wearing the camera out. The intense radiation and constant bombardment of particles are like a sandstorm hitting a delicate watch; over time, the wires get tired, some stop working entirely, and the pictures start to have missing pieces or blurry spots.

Scientists need to figure out how to draw the paths of these particles even when the camera is damaged. Traditionally, they use a step-by-step recipe: "First, find a dot here, then connect it to a dot there, then check if it fits a straight line." But when the camera is broken, the dots are missing, and the lines are jagged, so this old recipe fails. The big question is: Can we teach a computer to be smarter about connecting the dots when the picture is messy, without having to rebuild the whole camera? This is the challenge faced by the Belle II experiment, a massive physics project trying to understand why the universe is made of matter instead of antimatter.

The Broken Camera and the Smart Solver

In the world of the Belle II experiment, the main camera is a Central Drift Chamber (CDC), a giant cylinder filled with gas and thousands of wires. As the machine runs, the wires get "tired" from the radiation, losing their ability to detect particles. Some wires stop working completely, and others become unreliable. This creates a nightmare for the scientists: the data they get is full of holes. Imagine trying to trace a runner's path through a forest, but every few steps, a tree disappears, and sometimes a whole section of the forest is gone.

The old way of solving this, called the "Baseline Finder," is like a strict teacher who says, "If you miss two steps in a row, I don't believe you ran at all." It relies on finding a perfect, continuous line of hits. When the camera gets damaged, this teacher throws up their hands and stops counting, leading to a huge loss of data. In fact, when the detector was degraded in their simulations, this old method saw its efficiency drop by 28% compared to its performance on a healthy detector, and the tracks it did find were only 90% pure (meaning some were fake).

Enter the new hero of this story: a Graph Neural Network (GNN), which the authors call the "BAT Finder." Instead of a strict teacher, think of the BAT Finder as a super-smart detective who doesn't care about the order of the clues. It looks at all the scattered dots (hits) on the page at once and asks, "Which of these dots belong to the same story?" It uses a technique called "object condensation," which is like a magnet that pulls all the dots belonging to one particle together into a tight cluster, while pushing the noise and fake dots away. It doesn't need a perfect line; it just needs to see the pattern.

The Experiment: Training for the Worst-Case Scenario

The authors didn't just hope this detective would work; they put it to the test. They created a realistic simulation of a broken camera, mimicking the exact damage the real detector is facing: wires that are 35% less efficient (a value they call $C = 0.35$), entire sections of the detector that are completely dead, and huge gaps between different parts of the camera.

They tested three approaches:

1. The Old Way (Baseline Finder): The strict teacher.
2. The Middle Ground (CAT Finder): A mix where the detective only looks at the inner camera, but the teacher still checks the outer layers.
3. The Full Detective (BAT Finder): The detective looks at the whole picture at once.

First, they let the detective learn on a "perfect" camera. When they tested it on the broken camera without any extra training, it did better than the old teacher, but it still got confused by the big gaps, dropping its purity to about 93%. It was like a detective who is great at solving crimes in a clean city but gets lost when the streets are blocked.

But here is the magic trick: They retrained the detective. They fed it examples of the broken camera so it could learn what a "messy" path looks like. This didn't require building a new computer or changing the detective's brain; they just gave it more practice with the specific type of damage.

The Results: Saving the Day

The results were impressive. After retraining, the BAT Finder became a champion of the broken camera:

• Efficiency: It managed to find 64.2% of the uniformly displaced muons (specifically those displaced up to 100 cm). Compare this to the old method, which only found 36.2% of these specific particles. That means the new method saved nearly twice as many tracks for this group.
• Purity: Even with the messy data, the tracks it found were 96.1% pure. The old method only reached 90.1%.

The paper shows that the detective didn't just survive the damage; it adapted. When the camera had huge gaps where wires were missing, the old method failed completely because it couldn't bridge the gap. The BAT Finder, however, learned to connect the dots across those gaps because it looks at the whole event as one big puzzle rather than a sequence of steps.

Why This Matters

The authors conclude that the damage to the detector isn't a new, unsolvable problem; it's just a change in the "language" the data speaks. The patterns are different, but the physics is the same. By simply retraining the AI on the new, damaged patterns, they can keep the experiment running smoothly even as the hardware ages.

This is a big deal because the Belle II machine is going to run for years, and the detector will only get more damaged over time. Instead of panicking and trying to fix the hardware or rewrite the entire software from scratch every time a wire breaks, scientists can just hit "retrain" on their AI. It's a flexible, future-proof solution that ensures they won't lose the precious data needed to unlock the secrets of the universe, even as their camera slowly wears out. The paper proves that with the right AI, a broken camera can still take perfect pictures.