LHCb's GPU Trigger Uses Lipschitz Networks for Lepton ID

The LHCb collaboration has implemented new artificial intelligence algorithms to identify fundamental particles in real-time. The system uses Lipschitz-constrained neural networks running on graphics processing units to select events containing muons and electrons from proton collisions at the Large Hadron Collider.

According to a paper submitted to arXiv, the LHCb physics program in Run 3 critically depends on the efficient real-time selection of events with muons and electrons. These particles are key signatures in a wide range of heavy-flavor and exotic decay processes. The experiment's detector now operates with a fully software-based trigger that processes the complete detector readout at the LHC's bunch-crossing rate.

The Challenge of Real-Time Selection

In this environment, particle-identification algorithms must achieve high efficiency and background rejection while satisfying stringent constraints on throughput and memory footprint. The first trigger stage is executed on GPUs. This places unique computational demands on the software, requiring it to be both fast and accurate to avoid creating a data-processing bottleneck.

Lipschitz-Constrained Neural Networks

The researchers developed separate neural networks for muon and electron identification. These are Lipschitz-constrained networks, a type of AI model designed with mathematical properties that can improve stability and performance. The networks were trained using simulated collision events to recognize the subtle signatures of muons and electrons amidst a background of other particles.

The paper states that the performance of these new networks was evaluated relative to the previous baseline algorithms used by LHCb. The authors report that the new system demonstrates improved discrimination across a wide range of kinematic regions. Crucially, the improved performance is achieved while remaining compatible with the strict requirements of real-time GPU execution.

Performance and Application

The enhanced identification capability directly supports LHCb's core research goals. Muons and electrons are decay products of particles containing heavy quarks, like beauty and charm quarks. By more efficiently and accurately selecting events containing these leptons, physicists can better study rare decays, search for signs of physics beyond the Standard Model, and investigate charge-parity violation.

The successful deployment of these algorithms marks a significant step in integrating advanced machine learning techniques directly into the data acquisition chain of a major high-energy physics experiment. The work shows a trend in the field toward use GPU computing and specialized neural network architectures to handle the immense data rates of modern colliders. The implementation is now active as part of LHCb's Run 3 data-taking operations.