Use of Deep Learning techniques and Kinematic Fit with constraints in the search for the $H\rightarrow c\bar{c}$ channel at CMS

Two different analysis techniques to help improve the sensitivity of the direct search of the standard model Higgs boson decaying into a pair of charm quarks are presented. To reduce the amount of background in a direct $H\to c\bar{c}$ observation, the search is focused on events where the Higgs boson is produced in association with a $Z$ or $W$ boson, also called vector bosons. A total of five distinct final states are considered: $ZH\to \nu\nu c\bar{c}$, $WH\to e\nu c\bar{c}$, $WH\to \mu\nu c\bar{c}$, $ZH\to eec\bar{c}$, $ZH\to \mu\mu c\bar{c}$, divided into three different channels with 0, 1 or 2 charged leptons from the vector boson decay. A deep neural network is trained to perform a classification task to discriminate signal from background events. Its results are compared to the current machine learning method implemented in the analysis, a boosted decision tree with gradient boost. In each channel of the analysis different kinematic variables are used as the input, and different models are considered and tested for each channel. The area under the ROC curve shows a marginal improvement in each channel when a deep neural network is employed compared against a boosted decision tree. A least square kinematic fit is implemented in the $ZH\to l^+l^-c\bar{c}$ channel in order to achieve a better $m_{c\bar{c}}$ resolution. The low resolution in the c-jets momentum reconstruction can be improved by forcing some physical constraints on them. Conservation of momentum in the transverse plane and the assumption that the invariant mass of the two leptons should be equal to the mass of the $Z$ boson are the constraints exploited in this method. Up to a 30\% improvement in the mass resolution of the reconstructed Higgs boson was achieved by using this technique.