Master defense: Mads Juhl Storr-Hansen

Title: Improvement of Photon PID through Machine Learning

This thesis presents a tree-based Machine Learning method to improve the Particle Identification of photons in  and , which so far has been done by using a cut based method. In order to evaluate the results of the trained models several models have been trained. Initially single particle models for isolation and identification were trained for electrons, muons and photons, where these were improved compared to the ATLAS cuts by 19-22%, 1-15% and 40% respectively. 

These models were combined to particle pairs with  being improved by 67%,  by 13% and lastly  by 2%. Lastly, models for  and  were trained, which achieved 178% and 189% better results than ATLAS. These results seemed to be slightly overestimated, and will require additional testing.

Additional reconstruction was attempted, where the ATLAS cuts and the ML models were used on known MC data, for  and , and unknown real data for . Several different methods of applying the ML cuts were tested in MC, with the event with decays to electrons having an improvement of 19-128%, while the events with decays to muons had improvements of 8-137%, depending on the method used for selection. The diphoton decays obtained an improvement of 9-22%.

Testing on the real data in the  events resulted in improvements ranging from -5% to +2.5%. These results do not have any resemblance with the results achieved in MC data, thus the models would be required to be tested further, until they can achieve better results in both MC and real data.

Zoom link to this M.Sc. defense: