Electronics has never been smooth sailing in the field of nuclear physics. The Large Hadron Collider, the world’s most powerful accelerator, produces so much data that recording it all has never been a viable option.
As a result, systems that process signal waves from detectors are good at “forgetting” – they reconstruct the trajectories of secondary particles in less than a second and evaluate whether the collision they just observed can be ignored, or whether it is worth saving. for further analysis. However, current methods of reconstructing particle trajectories will soon no longer be sufficient.
Scientists at the Institute of Nuclear Physics of the Polish Academy of Sciences (IFJ PAN) have shown through research that tools built using artificial intelligence could be an effective alternative to current methods of quickly reconstructing particle trajectories. Their debut could come within the next two to three years, perhaps in the MUonE experiment supporting the search for new physics.
The research, titled ” Machine Learning based Reconstruction for the MUonE Experiment “, was published in ” Computer Science ” on March 10, 2024 .
Paper link: doi.org/10.7494/csc…
Significant progress has been made in the field of high energy physics (HEP) experiments over the past few decades, including computational technologies. The exploration of new physical phenomena is an extension of the so-called Standard Model, the current incomplete theoretical knowledge about the fundamental behavior of the fundamental components of nature and their interactions, leading to experimental studies at ever-increasing energies.
The number of particles produced by an interaction of two particles (a collision event) generally increases with the energy of the collision. Therefore, a large number of charged particles (such as in proton-proton collisions) must be reconstructed, leading to more complex patterns of events.
Illustration: An example of an event in a high-energy physics experiment, showing the trajectories of multiple particles through a detector. (Source: paper)
Particles collide in an accelerator to create a large cascade of secondary particles. Electronics that process the signal from the detector then have less than a second to evaluate whether an event is worth saving for later analysis.
In the near future, this difficult task may be accomplished using AI-based algorithms.
In modern high-energy physics experiments, particles emanating from a collision point pass through successive layers of a detector, depositing a bit of energy in each layer. In practice, this means that if a detector consists of ten layers, and a secondary particle passes through all of these layers, its path must be reconstructed based on ten points. The task seems simple.
“There is usually a magnetic field inside the detector. The charged particles move along a curve in it, and this is how the detector elements activated by them (called impacts) are positioned relative to each other,” explains Professor Marcin Kucharczyk from IFJ PAN.
“In practice, the so-called detector occupancy, i.e. the number of hits per detector element, can be very high, which causes many problems when trying to reconstruct particle trajectories correctly. In particular, reconstructing trajectories that are close to each other is a big problem The problem.”
Experiments aimed at finding new physics will collide particles with higher energies than before, meaning more secondary particles are produced with each collision. The brightness of the beam must also be higher, which in turn increases the number of collisions per unit time. In this case, classical methods of reconstructing particle trajectories are no longer adequate. AI excels in areas where certain common patterns need to be quickly recognized and can come to the rescue.
Deep neural networks for trajectory reconstruction
“The AI we designed is a deep neural network, including an input layer of 20 neurons, 4 hidden layers of 1,000 neurons each, and an output layer of 8 neurons. All neurons in each layer are connected. The network has a total of 2 million configuration parameters, the values of which are set during the learning process,” said IFJ PAN Dr. Milosz Zdybal.
Illustration: Neural network architecture for trajectory reconstruction. (Source: paper)
The resulting deep neural network was trained using 40,000 simulated particle collisions, supplemented with artificially generated noise. During the testing phase, only hit information is fed into the network. Because these come from computer simulations, the original trajectories of the responsible particles are known accurately and can be compared with the reconstructions provided by the AI. Based on this, the AI learns to correctly reconstruct particle trajectories.
Professor Kucharczyk emphasizes: “In our paper we show that deep neural networks trained on appropriately prepared databases are able to reconstruct secondary particle trajectories as accurately as classical algorithms. This is very important for the development of detection technologies. Although training A deep neural network is a long and computationally demanding process, but the trained network responds immediately since its accuracy is also satisfactory, so we can optimistically consider using it in real collision situations. “
MUonE Experiment
A proof-of-concept solution based on machine learning technology has been implemented and tested in the MUonE (MUon ON Electron Elastic Scattering) experiment, which seeks new physics in the area of muon anomalous magnetic moments. This examines interesting differences between measurements of a physical quantity related to muons, which are about 200 times as massive as electrons, and predictions from the Standard Model, the model used to describe the world of elementary particles.
Measurements at American accelerator center Fermilab show that the so-called muon anomalous magnetic moment differs with certainty by up to 4.2 standard deviations, or sigma, from predictions of the Standard Model. At the same time, the physics community generally agrees that a significance higher than 5 sigma (corresponding to 99.99995% certainty) is an acceptable value for declaring a discovery.
Illustration: Comparison of measurements of anomalous muon magnetic moments with predictions from the Standard Model. (Source: paper)
If the accuracy of Standard Model predictions can be improved, the importance of differences suggesting new physics could increase significantly. However, in order to better determine the muon’s anomalous magnetic moment, it is necessary to know a more precise parameter value, the hadron correction. Unfortunately, mathematical calculations of this parameter are not possible.
At this point, the role of the MUonE experiment becomes clear. Among them, the scientists intend to study the scattering of muons on electrons of low atomic number atoms, such as carbon or beryllium. The results will allow a more precise determination of certain physical parameters that directly depend on hadron corrections.
If all goes according to the physicists’ plan, hadron corrections determined in this way will increase confidence in measuring differences of up to 7 sigma between theoretical and measured values of the muon’s anomalous magnetic moment, a hitherto unknown factor in physics. Existence may become a reality.
The MUonE experiment could begin as early as next year at Europe’s CERN nuclear facility, but a target phase has been planned for 2027, when Krakow physicists may have a chance to see whether the artificial intelligence they created can be useful in reconstructing particle trajectories. Confirmation of its effectiveness under real experimental conditions could mark the beginning of a new era in particle detection technology.
Reference content: phys.org/news/2024-0…