Fast Machine Learning in the CMS Level-1 Trigger for the High-Luminosity LHC

The Large Hadron Collider at CERN will undergo an upgrade in 2027 to increase the integrated luminosity delivered to its associated experiments by a factor of 10 over its lifetime. The higher luminosity will allow CMS to broaden its physics programme with the Higgs boson self coupling being a key measurement that will be accessible with this upgrade. The upgrade will also increase the number of simultaneous proton-proton interactions (pileup) to between 140 and 200. To maintain physics selectivity in this high pileup the CMS Level-1 trigger will be upgraded to use charged particle tracks from the silicon tracker for the first time and more sophisticated algorithms, including those based on machine learning, will be implemented on an all-FPGA architecture. Level-1 track reconstruction is an essential part of the Level-1 trigger upgrade; without it the trigger would be unable to cope with the increased pileup and will lose sensitivity. Work on the final module of the reconstruction algorithm is presented. A track quality boosted decision tree was designed, implemented and also shown to outperform basic cuts on track fitting parameters and has been run in hardware tests. Demonstrations of reduced versions of the track reconstruction algorithm were performed along with integration tests with the wider trigger system. The global track trigger uses the reconstructed tracks to perform primary vertex finding. The location of the hard scatter in an event is essential for reducing background pileup contributions to energy sums and reduces the number of candidates being used in downstream algorithms. An end-to-end neural network-based approach to primary vertex finding and track-to-vertex association was developed and shown to outperform the baseline approach. This model was successfully implemented in firmware, running within the required latency and resource usage, and initial hardware testing results are shown. THe use of machine learning models in a changing environment such as the Level-1 trigger is an open topic in trigger development. Different strategies such as robust learning, uncertainty quantification, and continual learning are proposed and their suitability for the different time scales of the changing environment in CMS is explored. A study into the use of continual learning in the trigger demonstrates that machine learning models are at risk of performance degradation in a changing environment, but a continual learning approach can prevent this from happening and produces more robust models; an essential feature for algorithms which decide whether data is kept