This group is part of the Compact Muon Solenoid (CMS) Collaboration working with data from the CERN Large Hadron Collider (LHC). The principal focus on measurements of new physical phenomena that can be observed in the data they have collected at the very high particle collision energy available at the LHC, which began in 2010 and will continue into the 2030s. Current work includes the development of new detector technologies for the next stage of the LHC program, where the data rates are expected to increase by a factor five or more, the development of improved analysis code for the data the researchers are collecting, and the design of efficient code to analyze the data they expect to collect in the next phase of the project.
The group's data are a large number of "events" that are kept after a real-time selection process. Of the 1 GHz of interactions they store about one thousand events or ~1Mb at a rate of about 1 kHz. These events are stored for further processing and data analysis.
The Minnesota CMS group has been using MSI since the summer of 2020 in a program to improve the reconstruction code of electromagnetic showers in their ~75,000 channel crystal calorimeter. By using graph neural networks (GNNs) running on the MSI GPUs, the researchers have developed a new reconstruction method for the CMS Collaboration that significantly improved upon the previous method.
Based on the success of this model, the group is expanding to other data analysis applications of GNNs to the data. Currently they are exploring the reconstruction of pion events in much more complex detector, known as the high granularity calorimeter (HGCAL), which they are currently designing and building for installation in 2027. This program has already shown considerable promise, efficiently reconstructing hadronic (proton and pion) events in a prototype of the HGCAL detector which had ~30,000 detector channels. Using GNNs the researchers have improved considerably over traditional reconstruction methods. They are now developing a GNN for the event reconstruction in the final detector, where the background events will be very large, much larger than in the prorotype.
In a separate study the group is developing a GNN-based classification model to identify photons in high energy collisions where there is a large background of jets of particles produced by quarks.