Neuromorphic Computing in High Energy Physics
Bartłomiej Borzyszkowski · Zenodo (CERN European Organization for Nuclear Research) · 2020
At particle colliders, more data are produced than what the experiments can store for further analysis. This is why the incoming collisions are processed in real time by a so-called trigger system. At the Large Hadron Collider (LHC), trigger systems are structured as a combination of two hierarchical layers: the first trigger level (or L1 trigger) operates with latency of O(10 μec), selecting the ~100K events/sec to be passed to the second stage. The second stage, called high-level trigger, runs a faster and coarser version of the offline reconstruction, to select the final 1000 events to be stored for analysis. The L1 trigger selection algorithms are deployed as signal processing through electronic circuits, usually implemented though ASICs or FPGAs. The HLT runs on a CPU farm, supposed to be powered by GPUs or FPGAs in the future. Recently, an effort to deploy neural network on the L1 FPGAs was started, resulting in a Deep-learning- to-FPGA firmware deployment (HLS4ML) being developed specifically for low-latency inference. This is only the first step into an R&D program aiming to explore new computing architecture to process LHC data in real time. In this respect, we are interested to explore emerging technologies for fast inference. Neuromorphic chips are clearly part of this R&D program, being a natural environment to implement Spiking Neural Networks (SNNs). Neuromorphic computing is an interesting candidate for signal processing at the High-Luminosity LHC, the next stage of the LHC upgrade (scheduled to start in 2025). For HL-LHC, existing particle detectors will be upgraded, that will allow to take a time-sequence of snapshots for a given collision. This additional information will allow to separate the signal belonging to the interesting collision from those generated parasitic collisions occurring at the same time (in-time pileup) or before/after the interesting one (out-of-time pileup). By powering the LHC real-time processing with SNNs, one could be able to apply advance and accurate signal-to-noise discrimination algorithms in real time, without affecting the overall system latency beyond the given tolerance. We propose to investigate the potential of SNNs deployed on neuromorphic chips as a technological solution to increase the precision of the upgraded CMS detector for HL-LHC. This includes the characterization of a particle type (classification) based on the recorded features or image representation. These information can be used to solve a classic LHC problem and eventually lead to determine whether a particle belongs to the interesting collision or to one of the parasitic events. The study is based on simulations and real data collected during tests at particle beams, collected in the context of the upgrade studies for the CMS detector.