Identification of highly Lorentz-boosted heavy particles using graph neural networks and new mass decorrelation techniques

This note presents several new developments on machine learning (ML)-based identification of highly Lorentz-boosted heavy particles using jet substructure in CMS and their performance with the CMS Phase 1 detector. A new algorithm based on ParticleNet, a graph neural network using an unordered set of jet constituent particles as the input, has been developed and shows significantly improved performance. Two new methods have been investigated to decorrelate ML-based particle identification algorithms with the jet mass, one based on the Designing Decorrelated Taggers (DDT) approach, and the other by training on an artificial signal sample generated with a flat mass spectrum for the signal particle. The new methods are compared with the existing one based on adversarial training and show better performance.