FPGA Implementation of a CNN-Based Topological Trigger for HL-LHC.

Area of Science:

• High Energy Physics
• Computer Science
• Machine Learning

Background:

• Fast online event selection is crucial for data processing at hadron colliders.
• Convolutional Neural Networks (CNNs) offer powerful image classification capabilities.
• Field-Programmable Gate Arrays (FPGAs) provide hardware acceleration for real-time applications.

Purpose of the Study:

• To investigate the implementation of CNNs in programmable logic for fast online event selection.
• To optimize CNN architectures for hardware constraints and latency requirements of the High Luminosity Large Hadron Collider (HL-LHC).
• To evaluate the physics performance of CNN-based event selection using simulated data.

Main Methods:

• Studied CNNs using full event images for classification.
• Incorporated hardware-aware optimization of network architectures.
• Evaluated physics performance with simulated data, focusing on signal/background discrimination.
• Assessed implementation feasibility within current FPGA resources and HL-LHC trigger latency.

Main Results:

• Identified CNN models suitable for current FPGAs and HL-LHC trigger latency.
• A candidate CNN model demonstrated excellent signal/background discrimination for the HH(bbbb) channel in the CMS L1 trigger.
• Performance is sensitive to pile-up mitigation techniques before image generation.

Conclusions:

• CNNs are viable for hardware implementation in fast online event selection at hadron colliders.
• Optimized CNNs can meet the demanding requirements of the HL-LHC trigger system.
• The proposed approach shows significant potential for enhancing physics analyses at future colliders.