Speaker
Description
Charged-particle track reconstruction is an important part of modern collider experiments such as ATLAS and CMS that will face challenging conditions in the future High Luminosity phase of the LHC due to high pile-up. The increasing time and compute costs associated with the current tracking algorithm have spurred the development of machine learning (ML) alternatives to high degrees of success. However, both the traditional and ML algorithms indiscriminately consider all hits collected in a collision event, but only keep a small subset of reconstructed tracks relevant to physics afterward, leading by design to a substantial waste in computation. To address this issue, we present an approach to improve their computational performance by selectively reconstructing tracks using only hits from target particles.
Our approach is a light-weight neural network for point cloud segmentation based on random sampling and local feature aggregation. The architecture processes information encoded in both the individual hit features and its local environment, resulting richer hit representation. Using TrackML and Open-Data Detector data, we demonstrated the model performance on segmentation task, reducing up to 50% non-target, while keeping 99% target particle hits. Furthermore, we show that the model is capable of learning useful representation for other tasks, such as metric learning for graph construction, suggesting a possible approach to pre-training directly on detector hit data that might benefit downstream tasks. We conclude with measurements of computational speed, to demonstrate the feasibility of this method as a preprocessing step to track reconstruction.