First-hand information for everyone
The High‑Luminosity Large Hadron Collider (HL‑LHC) will produce many more simultaneous proton collisions than today. That makes it harder for experiments to trace charged particles back to where they started. The CMS collaboration has added a new step to a fast, parallel tracking algorithm called Line Segment Tracking (LST). This new step, called a quadruplet or T4, increases the radius at which CMS can reliably reconstruct tracks that start away from the collision point — from about 40 cm to about 60 cm in the tests reported here. By comparison, the legacy tracking approach without LST reached only around 8–10 cm for such displaced tracks.
LST builds track candidates using hits in the outer tracker only. It first groups two close hits in the same module into a “minidoublet,” then links nearby minidoublets into a short straight piece called a line segment. Line segments are linked into triplets (T3) and then into longer objects like quintuplets (T5). The new quadruplet (T4) is defined as a linked pair of triplets that share a line segment. T4s are built only from triplets that were not already used to make T5s or pixel‑linked objects, and only in specific detector regions. The T4s are added to the final list of track candidates that the rest of the reconstruction uses.
Because displaced tracks are rarer than prompt tracks that come from the collision point, the team used a small deep neural network (DNN) to separate real displaced T4s from prompt or fake ones. The DNN labels each T4 candidate as real prompt, real displaced, or fake. “Displaced” here means a simulated track that begins more than 10 cm away from the primary interaction point in the plane perpendicular to the beam. The DNN uses simple geometric features of the hits and the scores of the triplets that make up the T4. The selection thresholds were tuned to keep between 95% and 99% of real displaced T4s while rejecting many prompt and fake candidates.
The authors tested performance with two kinds of samples. In a simple muon gun sample without pileup (10 muons per event with production points spread up to 50 cm from the origin), adding LST greatly raised the displaced‑track efficiency compared to legacy tracking, and adding T4s doubled the efficiency starting at about 15 cm of radial displacement. In simulated proton‑collision events with pileup 200 (a realistic HL‑LHC condition), LST increased displaced acceptance from roughly 10 cm to about 40 cm, and the new T4s extended that acceptance to about 60 cm. Prompt‑track efficiency stayed at a similar level. The paper reports only minor increases in the fake and duplicate track rates when T4s are included. Efficiency was measured by matching reconstructed tracks to simulated ones when at least 75% of the hits came from the same simulated track.
There are important caveats. The performance studies reported here are limited to the first two iterations of an offline tracking sequence. Some tests used a simple muon sample without pileup, and others used simulated collisions with pileup 200; real experimental conditions may present additional challenges. The DNN selection and explicit cuts (for example a lower bound on the transverse impact parameter) were tuned for a balance between keeping displaced tracks and rejecting fakes; different tuning could change the tradeoffs. Finally, while the LST algorithm is designed to run efficiently on graphics processors and has been integrated into the CMS software, its impact on full, final workflows under real HL‑LHC running will depend on further tests and operational choices.