Machine Learning Improvements to the CMS Trigger System

The Large Hadron Collider at CERN is the world’s most powerful particle accelerator. Data from its proton-proton collisions is collected by the Compact Muon Solenoid detector (CMS) and is used to search for new physics phenomena, such as new particles that may be dark matter candidates. CMS produces hundreds of terabytes of data every second. It is impossible to store this much data, so the CMS trigger system vets out less interesting and valuable data. This system must make extremely fast decisions on which pieces of data to discard and keep. We have been able to improve the way the trigger system recognizes the trails of different types of particles (“jet tagging”) with machine learning, and can implement these improvements in a piece of hardware called a field- programmable gate array (FPGA). We continue to improve jet tagging with machine learning techniques. We optimize binary and ternary precision neural network models, synthesize them for integration into FPGAs, and measure their accuracy and efficiency. We show that binary and ternary jet tagging models are as accurate as floating point precision models, and they tend to perform with satisfactory speed and low resource usage in FPGAs.