Speaker
Description
Transformer-based jet taggers achieve state-of-the-art classification accuracy at the LHC but remain too computationally expensive for the FPGA-based Level-1 trigger systems planned for the HL-LHC upgrade. We present two complementary attention architectures, SAL-T and PHAT-JeT, that close this gap by introducing physics-motivated structural priors into the attention mechanism while preserving classification performance.
SAL-T employs a structured sparse attention pattern that exploits the local angular geometry of jet constituents in $(\eta, \phi)$ space. By restricting attention to physically meaningful neighborhoods rather than computing dense pairwise interactions, SAL-T substantially reduces FLOPs and parameter count while maintaining competitive performance on standard jet-tagging benchmarks.
PHAT-JeT replaces token-level self-attention with a hierarchical patch-cluster attention mechanism over the $(\eta, \phi)$ plane, motivated by patch-based vision transformers but adapted to the variable-multiplicity, ordering-robust nature of jet constituents. The hierarchical aggregation captures multi-scale jet substructure with significantly fewer operations than dense transformers, yielding a Pareto-optimal accuracy, latency, and resource frontier across model scales.
We evaluate both architectures on the JetClass dataset across ten jet categories and compare against established baselines including ParticleTransformer and ParticleNet. We further demonstrate hardware deployability via hls4ml synthesis, reporting resource utilization and latency on Xilinx FPGA targets relevant to the CMS L1 trigger upgrade.
Taken together, SAL-T and PHAT-JeT show that incorporating locality and hierarchy as architectural priors enables transformer-quality jet tagging within the strict latency and resource budgets of real-time trigger systems, contributing to the broader effort to deploy modern ML reconstruction directly in the LHC trigger pipeline at the HL-LHC.
| Do you plan to submit a 4-page extended abstract on OpenReview (only for Presentations/Posters)? | Yes |
|---|