September 10, 2025• Physics 18, 155
A deep neural network has proven essential in confirming a key prediction of one of the standard model’s cornerstones.
CERN
Figure 1: The eight toroid magnets of the ATLAS experiment bend charged particles in the inner detector and the muon spectrometer, enabling their electric charges and momenta to be measured.
CERN
Figure 1: The eight toroid magnets of the ATLAS experiment bend charged particles in the inner detector and the muon spectrometer, enabling their electric charges and momenta to be measured.×
The Higgs mechanism explains why the electromagnetic and weak interactions have such drastically different strengths—that is, how their symmetry became broken a picosecond after the big bang. The Higgs does not interact with photons, rendering them massless, whereas they do interact with the carriers of the weak interaction (the W+, W–, and Z bosons), giving them masses of order 100 GeV. Their nonzero masses allow them to acquire a longitudinal polarization—that is, a spin orientation perpendicular to their direction of motion. Because of special relativity, photons and other massless bosons that travel at the speed of light can’t have longitudinal polarization, but the W and Z bosons and other massive particles can. If electroweak symmetry had been broken not by the Higgs mechanism but by a different interaction, there would be no Higgs boson to find. However, theorists predicted in 1985 that an alternative to the Higgs mechanism called technicolor would show up in collisions at energies of 40 TeV and above [1]. In 1999, theorists made a more explicit prediction: A future linear collider of sufficiently high energy and luminosity would inevitably find either the Higgs boson or the telltale sign of technicolor [2]. In 2012 the ATLAS and CMS experiments at the Large Hadron Collider (LHC) settled the question by observing the Higgs boson in proton–proton collisions at 8 TeV. Now the ATLAS experiment has verified another essential aspect of Higgs physics by detecting longitudinally polarized photons emerging from proton–proton collisions at 13 TeV (Fig. 1) [3]. The observation opens the door to further exploration of electroweak symmetry breaking and to searches for physics beyond the standard model.
Detecting longitudinally polarized W bosons entails separating them from transversely polarized ones. It’s a daunting task. The initial state occurs when W bosons radiate off incoming quarks in the protons smashed together at the LHC. These bosons interact to create more bosons that subsequently decay into quarks or leptons. Kinematically, longitudinally polarized weak bosons behave very similarly to transversely polarized weak bosons, except for the angular separation between their decay products. As a result, it is extremely difficult to distinguish signals from other collisions at the LHC.
To mitigate the difficulty, the ATLAS Collaboration designed a deep neural network (DNN) that identifies collisions that spawn at least one longitudinally polarized W boson. The DNN was trained on simulations to accumulate tiny differences into a numerical value called the discriminator, which distinguishes between the longitudinally and transversely polarized collisions. The ATLAS Collaboration separated collisions into regions of varying signal-to-background ratios based on the discriminator and used state-of-the-art theoretical predictions to model the various backgrounds. The number of collisions in the various regions were estimated for the various backgrounds and simultaneously fitted to a multidimensional numerical model to predict the Poisson likelihood of the number of longitudinally polarized W bosons for each region.
The DNN-powered search uncovered an excess of collisions over the nonlongitudinally polarized background. The 3.3 𝜎 signal corresponds to a cross section for the process of 0.88 ±0.30 femtobarns, which agrees with the standard-model prediction.
This is a wonderful result that confirms, with a high degree of statistical confidence, the Higgs mechanism’s core prediction of longitudinally polarized W bosons. It also spotlights a hotly debated issue. The search distinguished signals from the backgrounds without using any substantively interpretable variables that humans could arrive at. What does one do when artificial intelligence (AI) is used without a “human readable” kinematic interpretation of the signal? Although the DNN approach is quite useful for confirming predictions that come with a reasonable expectation of being accurate, it raises questions as to exactly how far we scientists will trust these AI models to do more and more separation of signal from background without a human being able to say, “Aha! This is the one variable I can see that signifies the process I’m interested in.” We must train ourselves as much as we train these networks to understand what is being done.
Speculating further, a DNN created to distinguish physics beyond the standard model could observe a signal in some difficult-to-interpret region. Many of my colleagues are divided about how to handle these cases and about how far to trust the predictions of these models. Time will tell how this plays out. In the meantime, the ATLAS Collaboration has confirmed a long-standing prediction of one of the cornerstones of the standard model.
ReferencesM. S. Chanowtiz and M. K. Gaillard, “The TeV physics of strongly interacting W’s and Z’s,” Nucl. Phys. B 261, 379 (1985).J. R. Espinosa and J. F. Gunion, “No-lose theorem for Higgs boson searches at a future linear collider,” Phys. Rev. Lett. 82, 1084 (1999).G. Aad et al. (ATLAS Collaboration), “Evidence for longitudinally polarized W bosons in the electroweak production of same-sign W boson pairs in association with two jets in pp collisions at s= 13 TeV with the ATLAS detector,” Phys. Rev. Lett. 135, 111802 (2025).About the Author
Salvatore Rappoccio is a professor at the University at Buffalo, New York, Department of Physics, which he joined in 2012. Before then, he was a postdoctoral researcher at Johns Hopkins University in Maryland. He earned his PhD in physics at Harvard University. A member of the Compact Muon Solenoid (CMS) Collaboration, he is interested in searches for new physics, boosted jets, measurements involving the top quark, jet substructure, silicon pixel tracking detectors, and the hierarchy problem.
Subject AreasRelated Articles
Atomic and Molecular PhysicsEnvisioning a Neutrino LaserSeptember 8, 2025
A Bose-Einstein condensate of radioactive atoms could turn into a source of intense, coherent, and directional neutrino beams, according to a theoretical proposal. Read More »
