New Techniques for Reconstruction in the CMS High Granularity Calorimeter

The calorimeter endcaps of the CMS detector are to be completely replaced by a High Granularity Calorimeter (HGCAL), as part of the Phase-2 upgrades. A novel reconstruction framework, called The Iterative Clustering (TICL), is under development in CMS Software (CMSSW) and was recently updated to its fourth version (v4) -- centered around a new, density-based, parallelizable, algorithm for pattern recognition, CLUE3D. CLUE3D is a single blob shower finder, tuned for high pile-up rejection, meaning showers are often split into several reconstructed 3D objects, called ‘tracksters’, especially when there are secondary components like the ones initiated by Bremsstrahlung for electromagnetic objects or in hadronic showers. A new linking algorithm was developed and integrated as part of TICL v4 with the goal of accumulating together these tracksters coming from the same shower. It uses the propagation of tracks and tracksters to the same surface to find links geometrically, and then build objects intelligently with those links found. Excellent reconstruction efficiencies were obtained for electromagnetic objects using TICL v4. Separate components of hadronic showers were also observed to be successfully merged using the linking algorithm. Moreover, improved jet energy response and resolution were observed with the new version of TICL. A more advanced approach based on learning structures of showers on a sparse graph of nearby tracksters, using a Graph Neural Network, was also explored. The problem was framed as an edge classification task. Communities were identified in the similarity graph obtained as the model output using spectral clustering. Clustering results better than those obtained with the geometric algorithm have been observed for both events containing two close-by pions and ten randomly chosen particles shot in front of the HGCAL. This approach is complementary to the recent efforts using much lower-level information like hits, and framing the problem as a node classification task.