The Large Hadron Collider (LHC) at CERN is currently the world's largest particle accelerator, generating approximately one gigabyte (GB) of data per second. Even after compression, the data accumulated by the LHC in one hour is equivalent to the amount of data collected by Facebook in an entire year. This massive amount of data presents enormous challenges for storage and analysis. Fortunately, particle physicists don't have to process all this data themselves. They collaborate with artificial intelligence (AI) called machine learning to handle it.
Scientists from the Stanford Linear Accelerator Center (SLAC) and Fermilab, both under the U.S. Department of Energy, summarized the current applications and future prospects of machine learning in particle physics in an article published in Nature on August 2.
"Machine learning algorithms know how to perform various analyses on their own, which promises to save us countless hours of design and analysis work," said Alexander Radovic, a co-author of the paper and a professor at the College of William & Mary. Radovic is currently involved in the NuMI off-axis neutrino experiment (NOVA) at Fermilab.
Machine learning to screen big data
Machine learning has proven highly successful in the field of analytics. To handle the massive amounts of data generated in modern experiments like those conducted at the LHC, researchers use so-called “triggers”—specialized hardware and software that determine in real time which data can be saved for analysis and which can be discarded.
Mike Williams, one of the paper's authors from MIT, said that machine learning algorithms can make at least 70% of the decisions. Williams is currently involved in the LHCb experiment, which helps scientists reveal why there is far more matter than antimatter in the universe.
The massive ATLAS and CMS instruments at the LHC are used to detect the Higgs boson. Each detector has millions of sensing elements, and their signals need to be put together to get meaningful results. Michael Kagan of SLAC said, "These signals form a complex data space, and we need to understand the relationships between them to draw conclusions, such as whether the trajectory of a particle in the detector is produced by an electron, a photon, or something else."
Neutrino experiments also benefit from machine learning. NOVA studies how neutrinos transform from one type to another as they pass through the Earth. These neutrino oscillations may potentially reveal the existence of a new type of neutrino, which some theories suggest is a particle of dark matter. NOVA's detectors are monitoring charged particles produced when neutrinos collide with detector materials, and machine learning algorithms can identify them.
Identify features for simulation
Machine learning algorithms are becoming increasingly complex and sophisticated, opening up unprecedented opportunities to solve problems in particle physics. The latest development in machine learning—so-called deep learning, which uses neural networks—has improved the way particle physicists conduct experiments.
Kagan said that many of the new tasks they can use deep learning are related to computer vision, “It’s similar to facial recognition, except that in particle physics, image features are more abstract than ears and noses.”
Data generated by experiments like NOVA can be easily transformed into actual images, from which AI can easily identify features. Radovic said, "Even if the data doesn't look like an image, we can still use computer vision methods if we can process it the right way. One area where this method is very useful is analyzing the large particle jets produced by the Large Hadron Collider."
Another emerging application of deep learning is the simulation of particle physics data, such as predicting what will happen during particle collisions in the LHC and comparing the results with actual data. Traditional simulations are typically slow and require enormous computational power, while AI can perform simulations much faster.
Kagan said, "While this is very early work, it shows a lot of promise and could help address future data challenges."
Questioning promotes progress
Despite significant progress, machine learning enthusiasts often face skepticism from their partners, partly because machine learning algorithms are mostly like "black boxes," offering little information about how they arrive at a particular conclusion.
Williams believes: "Questioning is a good thing. If you're using machine learning as a trigger to discard data, like we did in LHCb, then you need to be very careful and set very high standards. Therefore, building machine learning in particle physics requires continuous effort to better understand the inner workings of algorithms and to cross-check them with real data as much as possible."
Kazuhiro Terao, a co-author of the paper and a SLAC researcher working at the MicroBooNE neutrino experiment, said: "In applying AI, we should keep trying and always evaluate the results. Questioning should not be an obstacle to our progress. Today we mainly use machine learning to find features in data. In 10 years, machine learning algorithms may be able to independently ask questions and identify them when new physics is discovered."
