Trackster ID and cleaning for EM iterations

The Iterative Clustering (TICL) is a modular framework developed for iteratively reconstructing particles in the high granularity calorimeter (HGCAL) which will be installed in the endcaps of CMS for Phase II of the LHC. This note reports on the development of algorithms which improve the purity of the electromagnetic (EM) iteration in TICL in the presence of pile up (PU). A novel machine learning (ML) based particle identification (PID) technique is presented which uses graph neural networks to exploit particle shower properties in order to distinguish between unconverted photons and early showering pions, so that these pions can be filtered out from the EM step. Finally, a cleaning algorithm removing PU contaminations from reconstructed EM clusters and improving cluster properties is presented.