Vertex reconstruction for a search for neutron-antineutron conversions with HIBEAM
This paper evaluates and compares classical and graph-neural-network-based methods for reconstructing annihilation vertices in the HIBEAM experiment, finding that while both approaches achieve comparable accuracy for vertex coordinates, machine learning offers unique advantages in extracting event-shape information for downstream analysis.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The universe is filled with matter, from the stars above to the atoms in our own bodies, yet it is a profound mystery why there is so much of it and so little of its opposite, antimatter. According to the laws of physics as we currently understand them, the Big Bang should have created equal amounts of both, which would have immediately annihilated each other, leaving nothing behind. To explain why we exist, scientists look for rare violations of a fundamental rule called baryon number conservation. This rule states that the total number of heavy particles, like protons and neutrons, should remain constant. If a neutron could spontaneously transform into an antineutron, it would break this rule and offer a clue to the cosmic imbalance. Such a transformation is incredibly difficult to catch because it has never been seen, and the signals it might produce are easily drowned out by background noise.
To hunt for this elusive event, a new generation of experiment is being planned at the European Spallation Source in Sweden. The plan involves sending a beam of cold neutrons through a shielded tube and into a thin foil. If a neutron turns into an antineutron, it would immediately crash into the foil and explode into a shower of charged particles. The key to identifying this explosion is to pinpoint exactly where it happened on the foil. The experiment uses a large detector filled with gas, called a time projection chamber, which records the paths of the particles as they fly out. However, the detector also sees other particles, such as electrons scattered by stray radiation, which can look very similar to the signal. The challenge is to sort the true explosion from the noise and find the center of the event with extreme precision.
A team of researchers recently tested how well different computer methods could solve this puzzle using detailed simulations of the HIBEAM detector. They created millions of virtual events where an antineutron annihilates on the foil, producing a few charged pions that travel into the detector. To make the test realistic, they also injected extra background tracks, simulating the random electrons that the detector would see in a real run. The researchers then ran four different reconstruction strategies on this data to see which one could best calculate the location of the explosion. One method used a classic statistical approach that has been used in physics for decades, while another relied on a purely geometric trick that ignored the concept of tracks entirely. Two other methods used modern machine learning, specifically a type of artificial intelligence called a graph neural network, which treats the detector hits as a connected web of information to find patterns.
The results showed that for the simple geometry of this specific detector, the old-fashioned statistical methods performed just as well as the new machine learning tools when it came to finding the exact coordinates of the event. In the absence of background noise, all the methods could locate the center of the explosion with a precision of a few millimeters. However, the methods behaved differently when the background noise increased. The purely geometric method and the machine learning model that looked at the whole event at once were very stable; their accuracy barely changed even when the number of background electrons was doubled or quadrupled. In contrast, the classic statistical method and a more complex hybrid machine learning chain began to struggle as the noise grew, producing a few very bad guesses that stretched the range of error significantly.
Despite the similar performance in finding the center, the machine learning methods offered something the others could not. They provided extra information about the shape of the event, such as how many particles were likely involved and a confidence score for each guess. This extra data is valuable because it helps scientists decide later on whether an event is a genuine discovery or just a fluke. The study concluded that while the classic methods are sufficient for the basic job of finding the vertex, the machine learning approaches provide a richer picture of the event. The researchers also found that if they relaxed the rules to accept events with only a single visible track, they could catch almost every event, but the location of the explosion would become much less certain. Ultimately, the work suggests that for this specific search, the best path forward may be to combine the reliability of traditional methods with the descriptive power of machine learning to ensure that if a neutron does turn into an antineutron, the experiment will not only find it but understand it perfectly.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.