Speaker
Description
In the Large Hadron Collider (LHC) at CERN, protons collide more than a million times per second. Pileup, which are interactions in the same or nearby proton bunch crossings in the accelerator, can be thought of as noise which affects many reconstructed physics variables such as the Jet Mass, Jet PT, and missing transverse momentum. This noise also results in worse resolution and as a consequence lower physics reconstruction performance. Furthermore, pileup is expected to increase by more than a hundred times in the timespan that the energy in the LHC is increased until 2029. Currently, an algorithm called PUPPI (Pileup Per Particle Identification) exists to mitigate pileup. However, recent machine learning developments can provide a power to be able to more effectively remove this noise. A proof of concept study that uses a semi-supervised graph neural network for particle level pileup mitigation was previously tested using CMS fast simulation data. The idea is to connect the training samples (labeled) and testing samples (unlabeled) as nodes in a graph using tracking and physics information. In addition, a dedicated masking technique was applied to reduce bias. The model was retrained using CMS full simulation which introduced more complexity in geometry and consequently graph neighbor construction. This increase in the complexity of geometry necessitated a more complex masking technique of particle labels. To use hyperparameter tuning in conjunction with these masking parameters, we introduce a bayesian optimization framework that aims to minimize the performance metric: $$\tiny\frac{ \sigma }{1 - | \mu |}$$ for validation datasets. An improved performance of the reconstructed physics variables is obtained. The current results from the Semi-Supervised Graph Neural Network outperform the baseline pileup mitigation algorithm PUPPI.
