Speakers
Description
Charged-particle tracking is a core reconstruction task in high-energy physics that directly impacts all reconstruction and physics studies. At the High-Luminosity Large Hadron Collider (HL-LHC), this must be done under much higher pile-up while preserving both tracking quality and computational efficiency. While existing graph-based approaches achieve physics performance comparable to state-of-the-art traditional rule-based algorithms, their end-to-end runtime is often dominated by costly graph construction and processing. On the other hand, prior transformer-based approaches avoid explicit graph processing yet still rely on auxiliary stages such as hit filtering or clustering, making them not fully optimized end-to-end. To address these limitations, this work develops HEPTv2, a single-stage, end-to-end efficient point transformer for charged-particle tracking. HEPTv2 couples a locality-aware point encoder with a sectorized track decoder, enabling prediction of final tracks in an end-to-end trainable pipeline. On the TrackML dataset, HEPTv2 achieves 98.6\% double-majority (DM) tracking efficiency at a 0.8\% fake rate, with about 15 ms inference latency and 0.4 GB peak memory per event on a single NVIDIA A100 GPU, both of which scale near-linearly to $5\times10^5$ hits. This absolute operating cost brings learning-based tracking within reach of potential online deployment under HL-LHC-like rates. HEPTv2 also reaches the best accuracy–latency trade-off among compared methods, improving DM by more than 4.5\% over the strongest prior transformer baseline and by 1.1–2.2 % over highly optimized graph-based pipelines, while reducing latency by $7\times$ and $38–52\times$, respectively.
| Do you plan to submit a 4-page extended abstract on OpenReview (only for Presentations/Posters)? | Maybe |
|---|