GPU-based track-finding for the J-PARC muon g-2/EDM experiment

TL;DR

This work tackles the challenge of rapidly reconstructing positron tracks in the J-PARC muon g-2/EDM experiment under substantial pileup. It introduces a GPU-based parallelization of a Hough-transform track-finding pipeline, with a two-stage binning strategy and time-window parallelism, to deliver a substantial speedup while maintaining track-finding efficiency. Results show speedups of about seven to eleven times over the CPU baseline depending on GPU architecture, with tracking efficiency around ninety to ninety-five percent and manageable ghost-hit contamination. The findings support moving toward a GPU-first workflow, including porting track fitting to the GPU for a fully accelerated end-to-end simulation and data-processing chain.

Abstract

The muon \textit{g-2}/EDM experiment at J-PARC is designed to precisely measure the muon's magnetic moment and electric dipole moment, driven by discrepancies between theory and previous experiments. A key challenge is the fast reconstruction of positron tracks from multiple muon decays within a short time span causing an event pileup. One of the aspects is the identification of individual positron tracks from the reconstructed hits, which is currently done using a hough-transform based approach. Results from simulation studies have shown expected results in terms of efficiency and accuracy of track reconstruction. However, the execution time for the entire analysis chain is prohibitively long to be deployed in the experiment. Specifically, preliminary estimations suggest a requirement of 40 $\times$ speedup of the track-finding routine. In this context, we explore a GPU-based solution to accelerate track-finding through parallel processing and present the implementation details and the results of our study for different pileup conditions. The results indicate that the GPU solution far exceeds our expectation in terms of execution speed without compromising on the reconstruction efficiency.

Figures (8)

• Figure 1: Schematic view of the muon g - 2/EDM experiment at J-PARC MLF. Surface muons with a momentum of about 27 MeV/c from the production target start their journey from the H-Line and are brought to rest with about 25 meV kinetic energy through the formation of muonium in a silica aerogel medium. The muons are then transported to the muon LINAC where they are re-accelerated to an energy of 212 MeV to be finally injected into the magnetic storage ring. The muons orbit in the magnetic field until their journey ends as they decay to positrons and neutrinos. These positrons enter the detector volume placed in the storage ring. Note that positrons with a momentum of about 150 MeV/c cross the central void region of detector creating split tracks. The positron detector volume and the top view with a typical muon orbit and decaying positron orbit are also shown.
• Figure 2: Standard processing order of simulation, digitization and track reconstruction. The typical computation time required for each process is also shown for processing $10^4$ muon decay events in a single CPU.
• Figure 3: An example muon decay event. Left: RecoHits in three-dimensional space showing the hits due to the positrons. Middle: projection of the track shown in the left onto the $\phi$-$z$ plane. Right: the corresponding Hough space in which each curve is generated from a single hit and the intersection of the curves, i.e, the most populated bin in Hough space corresponds to the red line in the $\phi$-$z$ plane.
• Figure 4: Steps in the hough transform for track seed identification, showing hits in the $\phi$–$z$ plane, Hough space histograms, and the final seed. (a) is the projection of positron tracks of a multi-muon decay (pileup) event in the $\phi-z$ plane (b) is the same information transformed to the hough-space with a coarse binning of 18 bins along $\theta$ and 100 bins along $\rho$ with the yellow bin regions having a high likelihood of a track. (c) is a fine binned version (10 bins along each axis) of (b) around the high likelihood zone identified in (b). (d) is the line obtained from the bin with the highest count in (c).
• Figure 5: (a) Generator level momentum distribution of decay positrons from Geant4 (b) The momentum distribution from decay positron tracks identified by the track-finding algorithm.