Machine Learning Regression For Muon Momentum in the Level-1 Trigger of the CMS Experiment

We report on the retraining of a machine learning algorithm used for the regression of muon momentum in preparation for Run 3 of the Large Hadron Collider (LHC). The Compact Muon Solenoid (CMS) is a general-purpose particle physics detector at the LHC in Geneva, Switzerland. With proton-proton collisions happening nearly 25 million times every second, the CMS detector cannot store data from every event. To effectively decide which events to store for analysis, the CMS detector uses a trigger system. The trigger system brings the data rate from one hundred terabytes of data every second to only one gigabyte per second at the Level-1 Trigger (L1). The L1 trigger does this by computing an on-the-fly processing of the event in firmware and uses this information to perform selections based on objects in the event and their momenta. One such algorithm used by the L1 trigger is called the Endcap Muon Track Finder (EMTF). EMTF uses hits from subdetectors to build curved tracks as the muons travel in the solenoid's magnetic field. EMTF then utilizes a machine learning algorithm called a Boosted Decision Tree (BDT) for assigning momenta to each track. For Run 3, EMTF retrained the BDT using new data formats and training parameters to improve the performance for high momentum muons.