Speaker
Description
To meet the extreme data demands of the High-Luminosity LHC, the LCH experiments are deploying enhanced trigger architectures built on next-generation hardware. This infrastructure is engineered to handle several tenths of Tb/s with ultra-low latency of O(μs). Leveraging this processing power, offline-style algorithms that were previously beyond the computational reach of the online trigger stage can now be deployed. As a result, targeting rare phenomena at the trigger level is becoming feasible. Long-lived particles (LLPs), for example, are predicted by many beyond standard model scenarios and produce displaced vertices, delayed signals, and unusual energy deposits that conventional triggers—optimized for prompt signatures—often miss [1, 2].
The INTREPID project aims to bridge this gap by developing intelligent trigger strategies that embed machine-learning algorithms and advanced pattern recognition into the ultra-fast hardware for application to the CMS Level-1 muon trigger and other future experiments. In doing so, it enables sensitivity to these elusive signatures already at the earliest stage of data acquisition [3].
We present recent progress toward a unified framework for muon reconstruction and identification operating directly on low-level hit information from the muon system at MHz rates (Figure 1). Rather than relying on predefined reference patterns, this approach leverages the full detector granularity through flexible, learnable algorithms, enabling both standard and unconventional muon topologies to be identified within the trigger constrains.
A key component is the detection and tagging of muon showers—localised bursts of hits in the barrel muon system, typically produced by highly energetic muons or by LLPs decays. Our approach monitors hit multiplicities in a sliding time window; when a configurable threshold is exceeded, a compact descriptor encoding timing, multiplicity, and spatial spread is recorded and correlated with reconstructed muon tracks from neighbouring stations, promoting consistent combinations to dedicated high-momentum muon-shower trigger candidates. Preliminary tests on simulated samples demonstrate that the algorithm correctly tags over half of genuine showers while rejecting a large fraction of spurious candidates, improving retention of high-momentum muons without increasing noise rates. Initial firmware implementations have shown promising performance while maintaining low occupancy [4].
In parallel, a Graph Neural Network–based approach to muon tracking is being developed. Detector hits are modelled as graph nodes with edges encoding geometric compatibility in η-ϕ space, and a message-passing architecture aggregates information over variable-size neighbourhoods, naturally handling sparse, irregular occupancy and accommodating displaced geometries relevant to LLPs searches.
An end-to-end workflow bridges PyTorch Geometric training with FPGA-oriented C++ implementations synthesised using Vitis HLS. Post-training quantization to an integer-only INT8 datapath with data-driven bit-width optimization preserves model accuracy within 0.1% of the floating-point baseline while reducing memory footprint and arithmetic complexity. Bit-exact agreement between Python integer emulation and HLS C-simulation has been demonstrated, and an automated design-space exploration framework is being used to evaluate candidate designs across algorithmic, numeric, and HLS parameters. Quantization-aware training and physics-driven datasets for displaced-muon reconstruction are under development, targeting the latency budget of the hardware trigger [5].
These developments demonstrate how combining domain-specific pattern recognition, machine learning, and hardware acceleration can extend the discovery reach in HL-LHC era, particularly for LLPs and other unconventional signatures.