Source-linked AI summary
Track finding and particle identification performance with particle physics foundation models
Shuhang Li, Haiwang Yu, Gabor Galgoczi, Yeonju Go, Yi Huang, David Park, Michael Peters, Christof Roland, Jin Huang, Yihui Ren, Joseph D. Osborn
TL;DR
High-luminosity experiments face increasingly challenging reconstruction as detector data rates and occupancies rise. This paper evaluates a pretrained FM on sPHENIX TPC spacepoints for tracking and PID, finding improved low-pT tracking relative to CA and comparable or better PID performance than the conventional baseline.
Problem
Rising luminosities and streaming readout increase detector occupancies, making high-performance reconstruction increasingly challenging.
Source: §I · p. 1Method
The study applies frozen FM4NPP spacepoint representations through supervised adapters for track finding and particle identification on simulated sPHENIX TPC data.
Source: §I · p. 1, §II · p. 2Results
FM improves low-pT track-finding efficiency over CA while maintaining high purity; PID is similar for kaons and protons and slightly better for pions (AUC 0.792 versus 0.774).
Source: §III · p. 3, §IV · p. 4Takeaways & Limitations
A shared frozen pretrained backbone supports multiple reconstruction tasks through downstream adapters, despite pretraining not explicitly targeting energy or ionization-energy-loss prediction.
Source: §IV · p. 4The FM was trained on data from one detector subsystem, and further studies across detectors, collision systems, and experimental conditions are needed to determine representation generalizability.
Source: §III · pp. 3–4
Abstract
from arXiv · showhide
Modern high energy nuclear and particle physics experiments produce large amounts of sparse data that must be reduced significantly in size before performing physics analysis. Recently, there has been an effort to evaluate Foundation Models as a potential tool for performing data analysis. These models have been shown to learn features in high dimensional datasets based on the multi-modal data they are trained on. To show their effectiveness at performing complex pattern recognition tasks, such as reconstructing charged particles in a magnetic field, we evaluated the performance of a Foundation Model based reconstruction on several downstream tasks with a dataset composed of spacepoints in a Time Projection Chamber. We show that the performance of the Foundation Model based data analysis significantly improves upon that of traditional methods, and furthermore can be adapted to learn additional tasks that are not necessarily related to the initial training objective.
I. INTRODUCTION
Rising detector data rates and occupancies make high-performance reconstruction increasingly challenging. This work evaluates whether FM4NPP can bring transferable, low-level representations to tracking tasks and support multiple downstream tasks.
I. INTRODUCTION
- Higher luminosities and streaming readout at sPHENIX, with comparable or greater demands expected at future facilities, increase detector occupancies and reconstruction challenges.Source: §I · p. 1
- Foundation models learn general representations through pretraining and can be adapted to downstream tasks; self-supervision suits settings with abundant unlabeled data and costly, uncertain simulated labels.Source: §I · p. 1
- Prior jet-focused FMs have been adapted to classification, generation, regression, reweighting, unfolding, and anomaly detection, with additional work on calorimeter simulation and cross-domain transfer [2–6] [7] [8].Source: §I · p. 1
- Most prior FM studies use high-level or partially reconstructed inputs, while the cited Cherenkov-detector proto-FM uses hit sequences already associated with one track [9].Source: §I · p. 1
- Tracking ML includes graph, transformer, and state-space approaches, but most models are optimized for a particular objective using task-specific labeled samples [10] [11] [12] [13] [14].Source: §I · p. 1
- FM4NPP is pretrained on sPHENIX TPC spacepoints and evaluated on several tracking-related downstream tasks against conventional reconstruction.The model is designed to learn transferable representations from low-level detector data through self-supervised learning.Source: §I · p. 1
II. METHODS
FM4NPP is pretrained on simulated sPHENIX TPC spacepoints, then adapted for track finding and particle identification. The methods compare tracking with CA-based seeding and PID with a dE/dx-based baseline using specified matching and scoring conventions.
II. METHODS
- The sPHENIX TPC has 48 readout layers spanning radii of 30–78 cm, and simulated pp events provide spacepoints as FM training inputs.The approximately 10 million collisions at √s = 200 GeV were generated with Pythia8 and simulated through the full sPHENIX geometry with Geant4.Source: §II · p. 2
- FM4NPP uses a Mamba-based backbone, hierarchically serialized voxelized spacepoints, and self-supervised k-Nearest-Neighbor prediction; frozen representations feed task-specific supervised adapters.The track-finding adapter uses a transformer decoder and two MLP heads; the PID adapter uses a transformer decoder and one MLP classifier.Source: §II · p. 2
- Truth matching uses a double-majority criterion requiring at least 50% of both reconstructed-track spacepoints and truth-particle spacepoints to correspond.This enforces one-to-one matching and penalizes tracks that recover only fragments of a truth particle.Source: §II · p. 2
- Track finding is compared with CA seeding, which forms spatial-neighbor triplets, prunes them into chains, and extends seeds with Kalman Filter propagation.CA seeds are typically shorter than 16 layers to limit the CPU combinatorial burden; both methods use the same truth-matching criterion.Source: §II · p. 2
- PID compares FM classification with GPR-fitted dE/dx bands for π/K/p using ROC curves in the challenging p ∈ [0.8, 1.2) GeV/c interval.Both methods use the same conventional tracks with at least 20 spacepoints matched to primary particles, identical three-hypothesis populations, and normalized track-level scores; a single ROC routine isolates the score as the curve difference.Source: §II · pp. 2–3
- The FM and CA-based seeding track-finding efficiency and purity are plotted as functions of pT.The efficiency caption reports that FM-based seeding surpasses CA at low pT by a significant margin.Source: Figure 1 · p. 3, Figure 2 · p. 3
III. RESULTS
FM-based reconstruction improves low-pT track finding over CA-based seeding while maintaining high purity, including for looping tracks. Its PID performance is comparable to GPR dE/dx, with slightly better pion discrimination; generalizability beyond the studied detector data remains to be established.
Track finding
- At low pT, FM-based reconstruction significantly outperforms CA-based seeding, while both methods are highly efficient at relatively high pT.The FM reconstructs tracks across the full pT spectrum, including looping tracks that challenge conventional algorithms.Source: §III · p. 3
Track purity
- FM purity remains nearly unity at pT ≲0.5, while CA seeding struggles to join consecutive loops.The FM also avoids duplicate tracks among low-pT loopers.Source: §III · p. 3
Future work
- The studied FM was trained on one detector subsystem, and further studies are needed to assess generalizability across detectors, collision systems, and experimental conditions.The paper also identifies more generalizable pretraining objectives, potentially including energy loss, as a direction for improving downstream-task performance.Source: §III · pp. 3–4
Particle identification
- AUC 0.792 versus 0.774 for pions and 0.684 versus 0.676 for kaons; both methods achieve proton AUC ≥0.998.Values compare FM with GPR dE/dx in the selected momentum interval; the FM is slightly better for pions and comparable for kaons.Source: §III · p. 3, Figure 3 · p. 4
IV. CONCLUSIONS
On simulated sPHENIX TPC data, FM-based tracking improves low-pT efficiency while maintaining high purity, and a frozen pretrained backbone also supports particle identification. PID is similar to GPR-based dE/dx for kaons and protons and slightly better for pions, despite no explicit energy-loss prediction in pretraining.
Tracking
- On simulated sPHENIX TPC data, FM-based tracking substantially improves low-pT efficiency over CA-based reconstruction while maintaining high purity across the studied momentum range.At higher pT, the methods achieve similarly high tracking efficiency and purity.Source: §IV · p. 4
Particle identification
- The FM-based PID method is similar to GPR-based dE/dx for kaons and protons and slightly better for pion discrimination.The pretraining objective did not explicitly involve energy or ionization-energy-loss prediction.Source: §IV · p. 4
Implications
- A shared pretrained backbone can support several reconstruction tasks through appropriate downstream adapters.The authors suggest this could simplify data analysis and extend the reach of measurements, particularly at low pT.Source: §IV · p. 4