Commissioning and Low Latency Operation of the Graph Neural Network Electromagnetic Calorimeter Trigger at the Belle II Experiment

Imagine the Belle II experiment as a high-speed particle racetrack where tiny, invisible racers (particles) zoom around at nearly the speed of light, crashing into each other billions of times a second. Every time they crash, they leave behind a chaotic spray of energy, like glitter exploding in a dark room. The scientists need to catch the most interesting explosions, but there's a problem: there are too many crashes to record them all. The data storage would fill up faster than you can say "supercomputer."

To solve this, the experiment uses a super-fast security guard system called a "trigger." This guard has to look at every single crash, decide in a split second if it's worth saving, and shout "YES!" or "NO!" before the next crash happens. If the guard is too slow, the whole racetrack has to stop.

For a long time, the guard for the "Electromagnetic Calorimeter" (the part of the detector that catches energy bursts) used a very traditional, rule-based method called the ICN-ETM. It was reliable, but it was a bit like a guard using a simple checklist. Recently, scientists tried a new, smarter guard using Graph Neural Networks (GNNs)—a type of artificial intelligence that's really good at spotting patterns in messy data. They built a prototype called the GNN-ETM.

The Big Problem:
The original AI guard was too slow. It took 3.168 microseconds (that's 3,168 nanoseconds) to make a decision. The racetrack's rules said the guard must decide in 1.067 microseconds or less. If the AI took longer, it couldn't actually be used to stop the data flow; it could only watch and take notes. It was like having a genius detective who solves crimes perfectly but takes a week to write the report—too late to catch the criminal.

The Solution: Speeding Up the AI
The team in this paper didn't just build a faster computer; they completely redesigned the AI's brain and the house it lives in to make it lightning-fast. They did three main things:

1. Shrinking the Brain (Model Compression): They simplified the AI's "brain" by removing a layer of complexity and teaching it to use smaller, simpler numbers (like switching from high-definition photos to quick sketches) without losing its ability to spot the important crashes.
2. Rearranging the Furniture (Floorplanning): The computer chip (FPGA) they used is like a giant city divided into three districts. The original design had the AI's thoughts traveling back and forth between these districts, causing traffic jams. The team manually rearranged the AI's components so everything stayed in one district, cutting down the travel time.
3. Changing the Tools (DSP Mapping): They stopped using the chip's heavy-duty "hard" math tools because they were actually causing delays in this specific setup. Instead, they used the chip's standard logic, which turned out to be faster for this specific job.

The Result:
After these changes, the new AI guard, now called the Sunset CaloClusterNet, made a decision in just 1.053 microseconds. This is a massive improvement, cutting the time by a factor of 3.01.

Did it work?
The team didn't just guess; they tested it rigorously.

• Simulations: They ran perfect, cycle-by-cycle computer simulations that showed the new design matched the old, trusted offline models exactly (bit-for-bit accuracy).
• Real-World Testing: They installed the system in the actual Belle II experiment in December 2025. They measured the speed in real-time during "beam runs" (when particles are colliding) and "cosmic runs" (when the machine is off, but natural space particles hit the detector).

The measurements confirmed that the system is fast enough to meet the 1.067 microsecond deadline. In fact, with a specific setup, they had a safety margin of 160 nanoseconds. Without that specific setup, the margin was a tiny 14 nanoseconds, but it still passed.

What does it do now?
The new AI guard is now fully operational with its hardware connections established to the main control room (the Global Decision Logic), where it sends its "trigger bits" (the yes/no flags) for monitoring. However, in this specific commissioning phase, the system is not yet used to make the final active trigger decisions that stop the data flow. Instead, it runs in parallel to the existing guard, allowing scientists to compare its performance and validate its reliability in real-world conditions. The team also built a dashboard that lets them watch the AI's performance in real-time, counting how many "yes" decisions it makes every second.

When they compared the new AI guard to the old rule-based guard:

• In quiet times (cosmic runs), they found almost the same number of interesting events.
• In busy times (beam runs), the AI guard sometimes found more clusters because it's better at splitting up messy energy bursts, but when they added a special "signal classifier" (a filter to ignore background noise), the AI guard actually became more selective than the old guard, reducing the number of "yes" decisions by about 100 Hz in quiet runs and significantly in busy runs.

The Bottom Line:
This paper proves that a Graph Neural Network can be squeezed into a super-fast, real-time trigger system for a particle collider. They turned a slow, offline-style AI into a live, low-latency guardian that fits within the strict 1.067 microsecond deadline. The system is now running, monitoring its own performance, and ready to help scientists decide which particle crashes are worth saving for future study, paving the way for it to eventually take an active role in the trigger decision.