Implementation of a fast machine learning algorithm for the ATLAS phase-2 muon trigger

Nowadays high-energy physics experiments generate large data volumes at a faster rate than ever before. The Large Hadron Collider (LHC) delivers collisions every 25 ns, whereas the ATLAS and CMS experiments have to deal with tens of terabytes of data produced each second. Unfortunately, this data volume cannot be read at the original collision rate, then each experiment operates an online reduction system called the trigger. Most data acquisition systems have limitations on what they can collect and store, requiring a multi-level trigger to throttle the data stream and save it at a more manageable rate.

The ATLAS and CMS experiments deploy a two-stage trigger system to scan each LHC event at the full 40 MHz collision rate. The first trigger level (L1) operates within only microseconds, executing a decision within approximately 10 µs. L1 trigger is implemented using high-speed electronics, consisting of integrated circuit (ASIC) chips and/or field programmable gate array (FPGA) on custom cards, with high-speed optical interconnects. Major improvements made in the last years on several fronts can allow boosting the performance of the L1 systems, bringing them not far from what is achieved by processing offline in software full granularity data. Flexible tools became available to develop complex Machine Learning models and deploy them on generic commercial FPGAs. 

The recent phase-1 upgrade of the ATLAS Muon System replaced the first station of the forward region with the New Small Wheel, bringing improved trigger and tracking capabilities, with fast-track reconstruction at a rate up to 15 kHz/cm2 and an angular resolution of less than 1 mrad. I will present a study for classification and muon track reconstruction based on a FPGA fast graph neural network (GNN) method to be implemented in the future phase-2 ATLAS L1 muon trigger.