applicationContext = Production

Belle II: Picking the best data cherries with neural networks

The Belle II detector records about 5,000 collision events per second. However, only a fraction of these are of interest for physical questions. In order to distinguish the real signals from unusable data, the experiment uses “triggers”. Since the end of February, Belle II has been working with an innovative trigger system based on neural networks. The trigger now detects signals that would have previously been ignored. Yet the self-learning system is only at the beginning. With further training, the scientists expect even more precise results.

The display shows one of the first signals delivered by the neural trigger. It wouldn’t have been detected by the systems in place so far. (Image: Belle II Collaboration)

At the SuperKEKB accelerator, electrons and positrons collide in the Belle II detector. In the collisions, the experiment looks for rare decay processes. These could answer the question of why there is more matter than antimatter in the universe, among others.

“In order to actually find the few relevant signals, you need quite a lot of data,” says Christian Kiesling, a scientist at the Max Planck Institute for Physics and one of the brains behind the new development. “This requires that the SuperKEKB achieve the highest possible collision rates (i.e., high luminosity). However, the downside is that the collisions produce a lot of background noise in addition to the few usable results.”

SuperKEKB collision rates are so large that even the most advanced data collection systems are unable to transfer all events to external storage for later analysis. “We therefore need a selection process that can distinguish signals from the background in real time: a trigger,” explains Kiesling. “It’s quite agile in the process: The Belle II trigger has only a few millionths of a second to make the right decision.”

The human brain as a model

This is where artificial intelligence comes in: a machine that reliably handles the recognition of complex patterns in the microsecond range. Belle II relies on artificial neural networks. The architecture and complex circuitry of these are based on the biological model of human brain cells; they are then implemented on modern hardware. “This allows them to master the daunting task of rapid and complex pattern recognition,” adds Kiesling.

In Belle II, the neural trigger is designed to correctly detect particle tracks in an area of a few centimeters. These tracks originate from events that occurred during a collision and which are therefore classified as a signal. Tracks outside the collision zone are most likely due to background events.

After the first experiences with the new system, Kiesling is optimistic: “After only a few days, we can see how precisely the trigger works,” says Kiesling. “This will allow us to evaluate many more real signals – which will hopefully provide us with new insights.”

Cherry picking with the trigger

How does a trigger work? An example: A conveyor belt transports cherries (the signal) but also non-edible objects such as marbles and pebbles (the background). “Continuing with the physical analogy, there are far fewer cherries on the belt than there are marbles and stones”, explains Kiesling. “The conveyor belt runs in an opaque tube. Only through a small opening half a meter long – the detector – can you see cherries, marbles, and stones.”

The belt runs at high speed (high luminosity), and pebbles, marbles, and cherries pass through the opening very quickly. A person standing at the opening must select the cherries, thereby recognizing and distinguishing the passing objects on the belt (trigger), reaching out a hand, grabbing the cherries, and taking them from the belt (data recording).

In the accelerator experiment, the time for cherry or rather signal picking is only a few microseconds. In technical trigger systems, the experimenter chooses a set of simple logical conditions for the trigger; these are derived from certain elements of the detector. In the Belle II experiment, the trigger task is complicated by the fact that not only cherries but also a number of other interesting "fruits" with complex patterns are sought by the detector.

How do neural networks work?

Neural networks work with input variables (e.g., sensor signals), which are fed into a layer of “neurons”. These signals stimulate the network, and the network’s response or decision is determined by how hardwired input and output neurons are. These are also referred to as weights. “These weights need to be trained – just as with neurons in the brain,” explains Kiesling. “It’s consistent with the learning process.”

In artificial neural networks, this learning process is represented in a mathematical sequence. The network is presented with the input data multiple times, and the weights are changed until the outputs give the desired result.

Back to cherries, marbles, and stones: “The patterns of these objects can be used to train the weights of the neural connections,” explains Kiesling. “Different stimuli such as shape, size, and color are offered until the network correctly identifies the items.” The inputs and outputs of artificial neural networks are real numbers. Even the weights are numbers. You rate the objects on the conveyor belt with a number close to “1” (= cherry) or close to “0” (= pebbles or marbles).