Introduction
The data collected at the Large Hadron collider enables tests of the Standard Model with unprecedented precision and searches for new particles, such as candidates for dark matter or heavy resonances, even in rare and complex final states.
The top quark is the heaviest particle known to date and is among the most central objects to collider physics today: top quarks are omnipresent in collider searches both as background and as signal. The top quark Yukawa coupling being close to unity makes top quarks the most interesting objects in many scenarios for physics beyond the Standard model and provides an excellent opportunity for understanding the nature of electroweak symmetry breaking.
Recently, the ATLAS experiment has published the first evidence for the production of four top quarks. Searches in the four top quark final state have significant discovery potential for new particles which interact strongly to the top sector, such as resonances predicted in Randall-Sundrum models with warped extra dimensions. The student will analyse four top quark signal models and develop new reconstruction algorithms to improve the signal acceptance. Background processes which can mimic a four-top signal and need to be estimated using widely employed techniques. Hence, the student will gain a broad experience in the methods used in LHC searches and the underlying physics.
The main challenge is in the reconstruction and identification of hypothetical heavy resonances which couple preferably to top quarks. These heavy particles decay predominantly to pairs of top quarks and are produced in assiciation with two additional top quarks, giving rise to a four-top-quark final state. The two top quarks which originate from the decay of the resonance must be distinguished from the two so-called spectator tops which originate from the associated production of the resonance. Doing so allows for computing the invariant mass of the resonance top quarks and thereby identify potential heavy particles as "bumps" (localised excesses) in the invariant mass distribution of the resonance top quark pair.
This is not easy! A top, being an elementary particle with mass close to that of a gold atom, decays almost instantly after its production to a W boson and a bottom quark. The W boson itself also decays, either hadronically to a light quark pair or leptonically to a lepton and a neutrino. You can imagine that a four-top quark final state can be quite messy to reconstruct!
Therefore, we will tackle the problem in several layers of complexity, using simulated data describing a hypothetical heavy particle coupling preferably to top quarks.
First, we will look at a simulated process in which only the particles taking part in the hard scattering process are included. The decays of the top quarks are neglected for the moment, so that all the messy top quark reconstruction doesn't need to concern us. In the beginning, we have a simple problem: there are four top quarks with their four-vectors. We need to construct a classifier which can label them as "resonance" or "spectator" top quark based on their kinematic properties.
Next, we will investigate a more realistic scenario in which also the decay of the top quarks to W bosons and b-quarks are included, as well as the subsequent decays of the W bosons. The final state no longer is "just" the four top quarks but instead a wild mixture of heavy and light mesons from the hadronisation of the quarks and some more hadrons or leptons from the W boson decays. Before we can tackle the labelling of the resonance and spectator top quarks, we first need to reconstruct those. In this context we will explore jets as a method of clustering simulated particles originating from the hadronisation and parton shower of quarks and the decay of W bosons, as well as conventional methods for reconstructing top quarks. As the next step, we investigate how the machine learning classifiers from the simplified problem can be applied to this more realistic example.
Finally (if time allows), we will also consider background processes. In reality, the majority of the LHC collisions produces events with so-called dijet processes or the production of top quark pairs. The four-top-quark final state is extremely rare in comparison. For the description of observed data both the signal process of interest and all background processes which mimick it need to be accounted for. We will test the performance of our reconstruction method both for simulated samples of signal and background events to assess its performance.
The methods we will encounter for the classification of the resonance top quarks range from
traditionally employed techniques which are based on human-designed individual observables
multivariate techniques such as Boosted Decision Trees for classification based on several human-designed observables
fully connected deep neural networks for classification and graph neural networks for node classification
For every level of complexity of the problem, we will explore the different reconstruction techniques in their increasing complexity: first starting with human-designed observables ("the traditional way"), then moving to machine learning classifiers and comparing their respective performance.
Last updated
