Towards Real-Time Machine Learning-Based Signal/Background Selection in the CMS Detector Using Quantized Neural Networks and Input Data Reduction

Abstract

The Large Hadron Collider (LHC) is being prepared for an extensive upgrade to boost its particle discovery potential. The new phase, High Luminosity LHC, will operate at a factor-of-five-increased luminosity (the number proportional to the rate of collisions). Consequently, such an increase in luminosity will result in enormous quantities of generated data that cannot be transmitted or stored with the currently available resources and time. However, the vast majority of the generated data consist of uninteresting data or pile-up data containing few interesting events or electromagnetic showers. High-Luminosity LHC detectors, including the Compact Muon Solenoid (CMS), will thus have to rely on innovative approaches like the proposed one to select interesting collision data. In charge of data reduction/selection at the early stages of data streaming is a level 1 trigger (L1T), a real-time event selection system. The final step of the L1T is a global trigger, which uses sub-system algorithms to make a final decision about signal acceptance/rejection within a decision time of around 12 microseconds. For one of these sub-system L1T algorithms, we propose using quantized neural network models deployed in targeted L1T devices, namely, field-programmable gate arrays (FPGA), as a classifier between electromagnetic and pile-up/quantum chromodynamics showers. The developed quantized neural network operates in an end-to-end manner using raw detector data to speed up the classification process. The proposed data reduction methods further decrease model size while retaining accuracy. The proposed approach was tested with simulated data (since the detector is still in the production stage) and took less than 1 microsecond, achieving real-time signal–background classification with a classification accuracy of 97.37% for 2-bit-only quantization and 97.44% for quantization augmented with the data reduction approach (compared to 98.61% for the full-precision, standard network).