Improving missing transverse momentum estimation with a deep neural network

Abstract

At hadron colliders, the net transverse momentum of particles that do not interact with the detector (missing transverse momentum, (p) over right arrow sub(miss) over T) is a crucial observable in many analyses. In the standard model, (p) over right arrow sub(miss) over T) originates from neutrinos. Many beyond-the-standard-model particles, such as dark matter candidates, are also expected to leave the experimental apparatus undetected. This paper presents a novel deep neural network based. (p) over right arrow sub(miss) over T) estimator, DeepMET, developed by the CMS Collaboration at the LHC. The DeepMET algorithm produces a weight for each reconstructed particle based on its properties. The estimator is based on the negative vector sum of the weighted transverse momenta of all reconstructed particles in an event. Compared with other estimators currently employed by CMS, DeepMET improves the (p) over right arrow sub(miss) over T) resolution by 10%-30%, shows improvement for a wide range of final states, is easier to train, and is more resilient against the effects of additional proton-proton interactions accompanying the collision of interest.