EMC  

New Analysis Methods


Bildquelle: © P. Weber, Doktorarbeit

To improve precision measurements of the Standard Model and to search for New Physics, our group also develops new analysis methods. We explore new ideas in statistical methods, machine learning and electronics in order to maximize our measurement and discovery potential as well as minimize our systematic uncertainties. 

What we are working on

Quark/Gluon tagging on trigger jets

The analysis of dijet events can give us important information about new possible models beyond the Standard Model such as dark matter. In some models the massive mediator can be observed as an excess in the invariant mass spectrum of the decay product. However, dijet searches at masses below 1 TeV are statistically limited by the bandwidth and storage limitations of the de- tector. The trigger-object-level analysis (TLA) allows the search for low-mass resonances down to an invariant mass of 450 GeV by recording and analysing only a part of the full event information. In our signal model, the massive mediator decays into two quarks, while the QCD background is dominated by gluon-gluon and quark-gluon final states. Therefore, tagging the flavour of the two final state jets can suppress the background compared to the signal. TLA only uses information about the online reconstructed jets and some jet structure variables based on calorimeter information and therefore no tracking information is available. We therefore study how to tag quark vs. gluon jets using calorimeter information only. To do this, we use boosted decision trees and other machine learning methods. 

 

Autoencoders and Higgs to Invisible signatures

The search for Higgs decays to invisible particles is powerful search for New Physics, since the Standard Model prediction for this process is close to zero. Therefore any observation of this process would indicate physics beyond the Standard Model. However, selecting such decays from background processes such as Z boson decays to neutrinos is challenging. Modern machine learning such as autoencoders offer attractive solutions in enhancing Higgs to invisible signatures. The key idea of the autoencoder is that it ’encodes’ normal events and then fails to reconstruct an anomalous event that it had not yet encountered. In this case, our ’normal’ events are Z boson decays, while Higgs decays to invisible particles are reconstructed as an ’anomaly’. The autoencoder is therefore optimised on the background process only and is agnostic to the exact details of the signal process. By passing calorimeter cluster or track information to the autoencode instead of object information such as jet 4-momenta, we can exploit simultaneously many different pieces of information about the system.  

 

Previous Analyses

Jet Intercalibration

For many physics measurements in ATLAS, precise knowledge of the jet energy scale (JES) and its uncertainty is important. One way to validate the JES correction and to determine the uncertainty, is the jet pseudorapidity intercalibration, which is an in-situ measurement and balances dijet events. We developed a new method to select an increased statistical data sample of dijet events. Contrary to standard in-situ methods in ATLAS, which only utilize fully efficient triggers, this new method, called the Trigger Combination Method (TCM), combines many different, single jet triggers. Each individual trigger is not necessarily fully efficient, but the entire set is. The study includes comparisons of the obtained calorimeter response with the responses obtained from standard methods as well as the responses from Monte Carlo predictions. Furthermore systematic studies have been carried out to estimate the total systematic uncertainty of the TCM method. The results of this work can be found here.

 
up