My thesis subject is the study of polarisation of electroweak gauge
bosons. The longitudinally polarised state is highly correlated to
Goldstone boson so before the electroweak symmetry breaking, thus giving
us a way to test the limit of the Standard Model prediction. The
challenge of this observation of simultaneously produced bosons, called
Vector Boson Scattering, is the very low cross section.
During my second year of my PhD I continue my work on the deep neural
network (DNN) optimisation for signal vs background discrimination as
well as polarisation state determination for EW-WZjj. I proved that we
are able to outperform the previous machine learning based method that
were present in our framework. I set up the fit with a subset of
systematic and we are able to compute the significance for the
observation of the joint polarisation EW-W0Z0jj as well as single
polarisation state.
In the context of my work in the ATLAS collaboration and being an
author, I did a qualification task for the Combine Performance group
$e/\gamma$. I defined working points (WP) using DNN for identification
(ID) of electron in the forward region ($|\eta|$ > 2.5) for Run-4 with
High-Luminosity LHC, increasing greatly the number of collisions per
bunch crossing (pile-up, $\mu$). Thanks to the extension of the tracker
in this area I benefit from this new variables accessible to ID the
electron as well as calibrate their $p_T$ due to mismatch with previous
methods. We used a decorrelation technique to have the input of the DNN
$p_T$-independent making the output less $p_T$ dependent. Since this
work was efficient, we ported this work to make WP as well for the
current Run-3 and with more data we did an interpolation to define them
as function of $\mu$.
On the side of my work in the ATLAS Collaboration, I am involved on a
new way of scanning hyper-parameter space in Beyond the Standard Model
using DNN in the context of Active Learning. In collaboration with Ugo
de Noyers and Björn Herrmann from LAPTh, we're scanning scotogenic models
that they work on and already scan using Monte-Carlo Markov Chain. In
this way we are comparing this new methods with the pre-established one.
