Ever since the 1970s, when the third generation of quarks and leptons began to emerge experimentally, physicists have asked if further generations await discovery. One of the first key results from the Large Electron–Positron Collider 30 years ago provided evidence to the contrary, showing that there are only three generations of neutrinos. The discovery of the Higgs boson in 2012 added a further wrinkle to the story: many theorists believe that the mass of the Higgs boson is unnaturally small if there are additional generations of quarks heavier than the top quark.
But a loophole arises if the new heavy quarks do not interact with the Higgs field in the same way as regular quarks. The search for new heavy fourth-generation quarks – denoted T – is therefore the subject of active research at the LHC today.
CMS researchers have recently completed a search for such “vector-like” quarks using a new machine-learning method that exploits special relativity in a novel way. If the new T quarks exist, they are expected to decay to a quark and a W, Z or Higgs boson. As top quarks and W/Z/H bosons decay themselves, production of a T quark–antiquark pair could lead to dozens of different final states. While most previous searches focused on a handful of channels at most, this new analysis is able to search for 126 different possibilities at once.
The key to classifying all the various final states is the ability to identify high-energy top quarks, Higgs bosons, and W and Z bosons that decay into jets of particles recorded by the detector. In the reference frame of the CMS detector, these particles produce wide jets that all look alike, but things look very different in a frame of reference in which the initial particle (a W, Z or H boson, or a top quark) is at rest. For example, in the centre-of-mass frame of a Higgs boson, it would appear as two well-collimated back-to-back jets of particles, whereas in the reference frame of the CMS detector the jets are no longer back-to-back and may indeed be difficult to identify as separate at all. This feature, based on special relativity, tells us how to distinguish “fat” jets originating from different initial particles.
Modern machine-learning techniques were used to train a deep neural-network classification algorithm using simulations of the expected particle decays. Several dozen properties of the jets were calculated in different hypothetical reference frames, and fed to the network, which classifies the original fat jets as coming from either top quarks, H, W or Z bosons, b quarks, light quarks, or gluons. Each event is then classified according to how many of each jet type there are in the event. The number of observed events in each category was then compared to the predicted background yield: an excess could indicate T-quark pair production.
CMS found no evidence for T-quark pair production in the 2016 data, and has excluded T-quark masses up to 1.4 TeV (figure 1). The collaboration is working on new ideas to improve the classification method and extend the search to higher masses using the four-times larger 2017 to 2018 dataset.
CMS Collaboration 2018 CMS-PAS-B2G-18-005.