GNN-based track reconstruction for MUonE experiment

In the vast landscape of modern physics, scientists are constantly searching for cracks in the foundation of their most successful theory, the Standard Model. This theory acts as a rulebook for how the universe works at its smallest scales, yet it leaves some profound questions unanswered. One of the most persistent mysteries involves the muon, a particle similar to the electron but much heavier. For decades, experiments have measured how this particle wobbles in a magnetic field, a property known as its anomalous magnetic moment. The measurements consistently show a tiny, stubborn difference from what the Standard Model predicts. This gap suggests that unknown forces or particles might be influencing the muon, hinting at a deeper layer of reality waiting to be discovered. However, to confirm this, physicists need to measure the effect with extreme precision, requiring them to calculate a specific contribution from the interaction of particles with the vacuum of space. This calculation is notoriously difficult to do with traditional math, so a new experiment called MUonE was proposed to measure it directly by observing how muons scatter off electrons.

The challenge for the MUonE experiment is not just the physics, but the sheer volume of data it must process. The experiment will fire a beam of muons at a target at a rate of 40 million times every second. At this speed, the computer systems must decide in a fraction of a second which collisions are interesting and which are just background noise. If the system is too slow, it will miss the rare events that could reveal new physics. If it is too aggressive in filtering, it might throw away the very signals scientists are looking for. The traditional way to sort through this data involves complex, step-by-step algorithms that try to piece together the paths of particles from the tiny signals left in silicon sensors. While accurate, these methods are computationally heavy and struggle to keep up with the relentless pace of the beam. To solve this, a team of researchers in Poland turned to a different kind of intelligence: a machine learning system designed to think like a network of connections rather than a linear list of instructions.

The researchers developed a new method using a Graph Neural Network, a type of artificial intelligence that excels at understanding relationships between objects. Instead of treating the data as a grid of numbers or a picture, the system views the particle detector as a map of points and lines. Each point represents a place where a particle hit a sensor, and the lines represent the possible paths connecting those hits. The network's job is to figure out which lines are real paths taken by particles and which are just random coincidences. In a traditional approach, the computer might try to build every possible path and then check if it makes sense, a process that becomes impossibly slow as the number of particles increases. The new system, however, looks at the entire web of connections at once, learning to recognize the specific shape of a valid particle track.

To test this idea, the team used simulated data that mimics the conditions of the MUonE experiment, including the specific arrangement of sensors and the behavior of muons and electrons. They trained the network on thousands of these simulated events, teaching it to distinguish between the true paths of particles and false connections. The results were striking. The network successfully identified the correct paths with an error rate of less than one percent, a level of accuracy that matches the best traditional methods. More importantly, it did so with incredible speed. While the old algorithms could process about 3,400 events per second on a powerful computer, the new neural network handled over 31,000 events per second. This represents a tenfold increase in speed, effectively transforming a bottleneck into a smooth flow.

Beyond just finding the paths, the researchers also showed that this system could tell the difference between the particles themselves. Because electrons are much lighter than muons, they bounce off at sharper angles when they collide. The network learned to spot these subtle differences in the scattering angles, correctly identifying whether a track belonged to an electron or a muon. While it made mistakes in about 12 percent of these identification cases, this level of performance is promising for a system that is still in its early stages. The ability to perform both track reconstruction and particle identification simultaneously suggests that this approach could replace several separate steps in the data analysis pipeline.

The significance of this work lies in its potential to change how high-energy physics experiments operate. By moving the complex task of sorting data from the offline stage, where scientists analyze results after the experiment is over, to the online stage, where decisions are made in real time, the MUonE experiment can capture a much larger and cleaner sample of the rare events it needs. The researchers emphasize that while their results are based on simulations, the method is ready to be tested on real data collected during the experiment's upcoming test runs. If the system performs as well on real-world data as it did in the simulation, it could provide the necessary speed and precision to finally resolve the mystery of the muon's wobble, potentially opening a window into the new physics that lies beyond our current understanding.