Mass regression of highly-boosted jets using graph neural networks

In this note a novel technique is presented, based on machine learning (ML), to reconstruct the mass of hadronically decaying highly Lorentz-boosted heavy particles with the CMS Phase 1 detector. The technique, commonly known as mass regression, is based on ParticleNet [1-3], a graph neural network using an unordered set of jet constituent particles as the input. Mass sculpting is prevented by using a training sample with an artificially flat spectrum of the heavy particle mass. Compared to the more traditional grooming algorithms such as soft drop [4-6], the mass regression technique displays a significant improvement in the jet mass scale and resolution.