ML for di-tau mass reconstruction in ATLAS

The existing methods to reconstruct the Higgs boson mass from its tau channel exploit either a partially reconstructed mass or the collinear approximation. Recently, the Missing Mass Calculator technique (MMC) overcame these limitations by minimising an event likelihood. In my BSc thesis (University of Milano) I discussed two alternative solutions that exploit machine learning techniques. These techniques are competitive with the MMC in terms of performance, evaluated in terms of the Area Under the Curve (AUC), and have the advantage of being faster to compute. Firstly, I reproduced the results achieved by the Simon Fraser University group with a Boosted Regression Tree (BRT) through the Di-TauMassSKL package. Secondly, I enhanced the package capabilities by exploiting a Deep Neural Network (DNN), improving on the BRT’s results and achieving state-of-the-art results compared to the parametric method MMC.