Noise Filtering Algorithm Based on Graph Neural Network for STCF Drift Chamber

Hang Zhou; Jin Zhang; Shuangbing Song; Teng Li; Xiaocong Ai; Xiaoqian Hu; Xiaoqian Jia; Xiaoshuai Qin; Xingtao Huang; Xueyao Zhang

REVIEW 3 major objections 5 minor 2 cited by

Noise Filtering Algorithm Based on Graph Neural Network for STCF Drift Chamber

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding a graph neural network noise filter before Hough transform track finding restores STCF reconstruction efficiency at low transverse momentum to background-free levels and cuts the fake rate from 8.7% to 1.7%.

desk verdict Useful and plausible GNN preprocessing step for STCF tracking; the efficiency recovery is likely real, but the reported numbers need a proper train/test separation before they can be taken at face value. read the letter →

arxiv 2507.09224 v1 pith:AT52GRRP submitted 2025-07-12 hep-ex physics.ins-det

Xiaoqian Jia , Xiaoshuai Qin , Teng Li , Xueyao Zhang , Xiaoqian Hu , Shuangbing Song , Hang Zhou , Xiaocong Ai

show 2 more authors

Jin Zhang Xingtao Huang

This is my paper

classification hep-exphysics.ins-det

keywords STCFdriftchambergraphneuralnetworknoisefilteringtrackreconstructionHoughtransformfakeratebeambackground

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a graph neural network can act as a practical preprocessing filter for the STCF drift chamber, cleaning background noise before the existing Hough-transform track finder runs. The motivation is that STCF's high luminosity creates so much background that low-momentum tracks, particularly below about 600 MeV/c, are reconstructed inefficiently and many fake tracks appear. On Monte Carlo events of $J/\psi \to \pi^0\pi^+\pi^-$, the authors report that the filter restores standard-background reconstruction efficiency to the background-free level and reduces the fake rate from 8.7% to 1.7%, and from 21.6% to 2.5% at double background. If this holds in the real detector, the algorithm would be a practical upgrade to STCF's tracking chain without replacing the Hough finder.

What carries the argument

The load-bearing object is the graph neural network edge classifier together with the hierarchical threshold that turns its outputs into hit labels. Graph construction uses per-layer angle thresholds (1.24 rad for most layers, 1.5 rad at the U/V superlayer boundaries) to connect hits in neighboring layers, with connections skipping an empty layer to preserve track continuity and same-layer edges between adjacent hits. Node features are the wire position in $(r,\phi)$, the charge collected in the drift cell (ADC), and the recorded drift time (TDC). After eight graph-module iterations, each edge receives a weight interpreted as the probability that its two nodes belong to the same track. The tiered threshold applies no filtering at boundary layers 0 and 47, a low threshold of 0.1 in superlayers 0-1, and a high threshold of 0.5 elsewhere; a node is labeled signal if any of its edges passes, yielding 98.2% signal retention and 86.8% noise rejection.

What would settle it

Run the GNF-plus-Hough chain on a drift-chamber dataset whose background hits are not concentrated in superlayers 0, 1, and 4 at a mean of about 220 hits per event, for example data from a different beam condition or an alternative background simulation, and check whether the low-$p_T$ efficiency stays at the background-free level and the fake rate remains near 1.7%; a return to pre-filter efficiency or a large rise in fake rate would falsify the claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a GNN-based preprocessing stage can solve the drift chamber's noise problem without replacing the existing track finder. The GNF algorithm converts drift-chamber hits into a graph, where nodes are hits and edges are geometrically plausible connections between adjacent layers, with same-layer and skip-layer edges to preserve track continuity. An edge-classifying GNN is trained to output the probability that each edge connects two hits from the same track, and a tiered threshold converts those edge probabilities into signal and noise labels, deliberately leaving boundary layers unfiltered. On simulated $J/\psi \to \pi^0\pi^+\pi^- \to \gamma\gamma\pi^+\pi^-$ events, the authors find the filter keeps reconstruction efficiency at the background-free level under standard background while cutting the fake rate from 8.7% to 1.7% and, at double background, from 21.6% to 2.5%.

Load-bearing premise

The whole evaluation rests on Monte Carlo simulation with a specific background model: if the real STCF drift chamber's noise occupancy or spatial distribution differs from the roughly 220 hits per event concentrated in superlayers 0, 1, and 4, the tuned thresholds and trained network may not reproduce the reported gains.

Editorial extensions

If this is right

With GNF preprocessing, track reconstruction efficiency under standard background becomes comparable to the background-free case, including the previously degraded low-$p_T$ ($<600$ MeV/c) and large-$|\cos\theta|$ regions.
The event-level fake rate falls from 8.7% to 1.7% at standard background (80.5% relative reduction) and from 21.6% to 2.5% at double background (88.4% relative reduction).
The filter retains 98.2% of true signal hits and rejects 86.8% of noise hits, so the Hough transform receives a much cleaner hit set with almost no loss of genuine hits.
Track parameter resolutions ($d_0$, $z_0$, and $\sigma_{p_T}/p_T$) are unchanged by the noise filter, so the efficiency and purity gains come without degrading track quality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The filter's thresholds and network weights are tuned to one simulated background pattern; a natural robustness test would be training on a range of background occupancies and spatial concentrations so the gains survive a realistic mismatch.
Because the GNF stage is independent of the downstream finder, the same edge classifier could precede a Kalman-filter-based track finder or be folded into an end-to-end GNN tracker, not just the Hough chain.
The reported fake-rate numbers are event-level; if the goal is track-level purity, the next natural check is how many individual reconstructed tracks are fake before and after filtering, since an event with one fake track counts the same as an event with many.
A test-beam or prototype drift-chamber run would give the first check of whether the simulation's noise-hit distribution matches reality; if it does not, the tiered thresholds would need recalibration.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick read on arXiv:2507.09224. The GNF algorithm is a sensible preprocessing stage for the STCF Hough tracker. It builds a graph of MDC hits using candidate regions and skip-layer edges, trains an edge-classifying GNN, and maps edge scores to hit-level filtering via tiered thresholds. The MC study on J/psi events shows low-pT efficiency recovering to background-free levels and fake rate dropping from 8.7% to 1.7% at standard background, and from 21.6% to 2.5% at double background. The qualitative improvement is credible; the physics is well motivated and the comparison against the Hough-only baseline on the same data is a proper benchmark. The graph construction is tailored to the STCF geometry, and the tiered threshold idea is a simple but effective engineering choice that avoids killing boundary-layer hits.

The main soft spot is the evaluation protocol. The paper never states a train/validation/test split for the GNN. Section 4.2 says the thresholds were chosen by 'parameter space scanning,' and the same 200k-event datasets are then used to report the headline numbers. If the thresholds, the GNN hyperparameters (including the eight iterations), and the model itself were all tuned in-sample, the quoted AUC=0.999 and fake-rate reductions are optimistic. This is a legitimate concern, not a fatal one—the algorithm clearly separates true and fake edges in the score distributions, and the efficiency recovery is consistent across pT and cos theta, which is hard to fake entirely. But the authors need to add a held-out test set, report error bars, and describe the training data. They should also say whether the 200k events used for evaluation are disjoint from the GNN training sample.

Minor issues: no code or data release, which is normal for JINST but would help; no comparison with a simpler classifier, though the Hough baseline is the right reference for this study. The assumption that the OSCAR background simulation is faithful is stated but not validated, but that is standard at this stage.

Overall: this is a useful within-subfield contribution, and it should go to peer review. The core idea is sound; the missing train/test separation and in-sample thresholds are fixable with a clear re-analysis. Send it to referees.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes the GNF algorithm, a graph-neural-network-based noise filter intended as a preprocessing stage for STCF drift chamber track reconstruction. Hits are converted into a graph using geometry-based candidate regions, an edge-classifying GNN assigns connection probabilities, and a tiered threshold strategy maps edge predictions to signal/noise hit labels. The filtered hits are then fed to the existing Hough-transform track finder. Using Geant4/OSCAR Monte Carlo samples of J/psi -> pi0 pi+ pi- -> gamma gamma pi+ pi- events at two background levels, the authors report that GNF recovers track reconstruction efficiency in the low-pT region to a level comparable to the background-free case and reduces the fake rate from 8.7% to 1.7% at standard background and from 21.6% to 2.5% at double background.

Significance. If the reported gains hold on independent data, the GNF algorithm is a practical and well-motivated contribution to STCF tracking, where low-pT reconstruction in the presence of beam background is a known bottleneck. The graph-construction scheme is concrete, the two-background-level evaluation is sensible, and the comparison against the baseline Hough-only chain is a useful benchmark. The method is clearly described and the authors are appropriately cautious in calling the results preliminary. However, the central quantitative claims are not yet established as predictive because the manuscript does not document a train/validation/test separation for either the GNN or the threshold scan, so the reported efficiency and fake-rate numbers may be in-sample optima. This issue is fixable and does not invalidate the approach, but it must be addressed before the results can be taken at face value.

major comments (3)

[§4.2 and §4.3] The manuscript does not state whether the 200k simulated events used to report track reconstruction efficiency and fake rate are the same events used to train the GNN or to scan the tiered thresholds. Section 4.2 says the low/high thresholds (0.1 and 0.5) were selected by 'parameter space scanning' and then reports noise rejection of 86.8% and signal efficiency of 98.2%; Section 4.3 reports the combined GNF+Hough efficiencies and the 8.7%->1.7% and 21.6%->2.5% fake-rate reductions across 200k events. If the threshold scan and the evaluation used the same sample, all these figures are in-sample and the central claim that GNF restores efficiency to the background-free level is optimistic. Please specify a three-way split (training, validation, test), state the size of each split, and report all headline metrics on the test set only, including the criterion used during threshold selection.
[§4.3, Figures 8 and 9] No uncertainties are reported for the track reconstruction efficiency or the fake rate. Statements such as 'comparable to the case without background' and the 80.5% and 88.4% relative fake-rate reductions therefore cannot be quantitatively assessed, especially in low-pT and high-|cosθ| bins where the number of tracks per bin is smaller. At minimum, the authors should provide binomial confidence intervals (or another well-defined uncertainty) for each efficiency and fake-rate point and state the number of events or tracks entering each bin.
[§4.1, Figure 5] The entire validation rests on the assumption that the OSCAR 2.6.0 background simulation, with a mean of about 220 noise hits per event concentrated in superlayers 0, 1, and 4, faithfully represents the real STCF drift chamber environment. The paper provides no cross-check with alternative background models, no variation of the noise-hit rate, and no discussion of how the tuned thresholds or the trained GNN would behave if the true background composition or occupancy differs. The authors should either present a sensitivity study over background rate/composition or explicitly discuss this limitation as a caveat on the transferability of the reported numbers.

minor comments (5)

[Figure 7] The caption reads 'Architecture of Graph Neural Network' but the figure actually shows the full processing chain (MDC hit -> graph construction -> edge classification -> Hough transform -> GENFIT2); the caption should be corrected to describe the overall workflow.
[§3.1] The text '105 events/particle' should be written as 10^5 events/particle to avoid ambiguity about the statistics of the single-particle samples used to determine the angle thresholds.
[§4.2] The tiered-threshold description is imprecise: 'a low threshold of 0.1 in superlayer 0-1' should specify whether this means superlayers 0 and 1, and the exact superlayer or layer range for the high-threshold region should be stated explicitly.
[§3.2] The training setup is not described in enough detail for reproducibility: number of training events, number of epochs, batch size, learning rate, and model parameter count should be given.
[Figure 6] The caption says 'Training results of the Graph Neural Network'; if the ROC curve and score distributions shown are computed on the training set, this should be stated explicitly, and validation-set curves should be shown as well.

Circularity Check

1 steps flagged · score 4.0 of 10

Threshold operating points are selected by scanning on the same MC sample used for the headline efficiency/fake-rate numbers, so the reported gains are in-sample optima rather than held-out predictions.

fitted input called prediction [Section 4.2 'Signal hit selection and noise suppression' and Section 4.3 'Performance of the combined GNF-Hough Algorithm']
"Through parameter space scanning, the tiered threshold strategy configured with a low threshold of 0.1 in superlayer 0-1 and a high threshold of 0.5 in other layers achieves optimal track reconstruction performance. This configuration achieves a noise rejection rate of 86.8% ... while maintaining 98.2% signal selection efficiency. ... Metrics including track reconstruction efficiency, fake rate and spatial resolution are quantified across 200k simulated events."

The thresholds are chosen by scanning to optimize the very metrics that are then reported as the algorithm's performance (noise rejection, signal efficiency, reconstruction efficiency, fake rate). The paper does not state that the 200k events used for the final metrics are distinct from the data used for threshold scanning or for GNN hyperparameter selection ('eight iterations achieved optimal results'). The quoted headline numbers—86.8% noise rejection, 98.2% signal efficiency, and the fake-rate reductions to 1.7% and 2.5%—are therefore in-sample optima of the tuned operating point, not out-of-sample predictions.

full rationale

The paper's core comparison (GNF+Hough versus Hough on the same OSCAR/Geant4 MC events) is a legitimate, self-contained simulation study and is not circular in itself. The graph construction, GNN edge classification, and tiered-threshold mapping are all implemented from stated detector geometry and simulated data, with no load-bearing self-citation: refs. [13] and [3] are used only as prior work or motivation, not as evidence for the present result. The circularity concern is limited to the evaluation protocol: Section 4.2 says the tiered thresholds were obtained by 'parameter space scanning' to achieve 'optimal track reconstruction performance,' and Section 4.3 then quotes the efficiency and fake-rate metrics 'across 200k simulated events' without any indication of a train/validation/test split or a held-out sample for threshold selection. The reported AUC, noise-rejection rate, signal-selection efficiency, and fake-rate reductions are thus best understood as in-sample figures for the tuned operating point; they would not be expected to generalize to the same degree on independent data. This does not make the algorithm's advantage over the Hough baseline tautological, but it does mean the strongest quantitative claims are partly fitted rather than predicted. Score 4 reflects one instance of fitted-input-called-prediction in an otherwise non-circular, benchmark-based study.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physics entities. It relies on standard assumptions about detector simulation quality, the completeness of the graph construction, and the integrity of the training and evaluation split. The main free parameters are the graph angle thresholds and the tiered classification thresholds, all tuned on simulated data.

free parameters (4)

angle_threshold_for_graph_construction = 1.24 rad for most layer pairs, 1.5 rad at U/V superlayer boundaries
Determined from single-particle MC samples as the observed maximum deflection angle between adjacent layers; controls which hits are connected as candidate edges.
low_threshold = 0.1
Applied to edges in superlayers 0-1; selected by parameter scanning to maximize reconstruction performance on MC test data.
high_threshold = 0.5
Applied to edges in other superlayers; selected by parameter scanning on MC test data.
gnn_iterations = 8
Number of graph module iterations; chosen because tests show that eight iterations achieved optimal results (Section 3.2).

assumptions (4)

domain assumption Geant4/OSCAR simulation faithfully reproduces STCF drift chamber hit and background properties.
All performance evaluations use MC; if background model or detector response is inaccurate, the results do not transfer to real data (Sections 4.1).
domain assumption The angle thresholds are large enough that all true track segments are included as edges in the graph.
Graph construction must preserve complete connectivity of true tracks; the paper asserts the thresholds based on maximum deflection angles but does not prove there is no loss (Section 3.1).
domain assumption Training and evaluation datasets for the GNN are statistically independent.
The GNN is trained on some MC samples, but the training set is not explicitly described; the thresholds are scanned on the test datasets, so independence of the tuning step is not established (Sections 3.2, 4.2).
domain assumption Hough-based tracking, as implemented in OSCAR, is a valid baseline.
The comparison assumes the Hough algorithm performance is representative of the current STCF track finding; deficiencies in the baseline would affect the measured gain (Sections 2 and 4.3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Noise Filtering Algorithm Based on Graph Neural Network for STCF Drift Chamber." pith.science (2026). https://pith.science/paper/AT52GRRP

@misc{pith2026250709224,
  author       = {Pith},
  title        = {Pith review of: Noise Filtering Algorithm Based on Graph Neural Network for STCF Drift Chamber},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AT52GRRP}},
  note         = {Machine review of arXiv:2507.09224}
}

abstract

The super $\tau$-charm facility (STCF) is a next-generation electron-positron collider with high luminosity proposed in China. The higher luminosity leads to increased background level, posing significant challenges for track reconstruction of charged particles. Particularly in the low transverse momentum region, the current track reconstruction algorithm is notably affected by background, resulting in suboptimal reconstruction efficiency and a high fake rate. To address this challenge, we propose a Graph Neural Network (GNN)-based noise filtering algorithm (GNF Algorithm) as a preprocessing step for the track reconstruction. The GNF Algorithm introduces a novel method to convert detector data into graphs and applies a tiered threshold strategy to map GNN-based edge classification results onto signal-noise separation. The study based on Monte Carlo (MC) data shows that with the implementation of the GNF Algorithm, the reconstruction efficiency with the standard background is comparable to the case without background, while the fake rate is significantly reduced. Thus, GNF Algorithm provides essential support for the STCF tracking software.

Figures

Figures reproduced from arXiv: 2507.09224 by the authors.

**Figure 3.1.** Momentum distributions of charged particles from various processes at the truth level, nor Figure 1. Momentum distributions of charged particles from various benchmark physics processes in STCF. [PITH_FULL_IMAGE:figures/full_fig_p003_3_1.png] view at source ↗

**Figure 3.12.** Schematic layout of the STCF detector concept Figure 2Schematic layout of the STCF detector [PITH_FULL_IMAGE:figures/full_fig_p004_3_12.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Long-lived sterile neutrinos from axionlike particles at the Super Tau-Charm Facility
hep-ph 2026-05 conditional novelty 5.0 of 10

A proposed Chinese charm factory could detect heavy neutral leptons from axionlike-particle decays with a displaced-vertex search, reaching electron-neutrino mixing one to two orders of magnitude below current bounds.
Long-lived sterile neutrinos from axionlike particles at the Super Tau-Charm Facility
hep-ph 2026-05 unverdicted novelty 4.0 of 10

STCF can reach |V_eN|^2 values one to two orders of magnitude below current bounds for heavy neutral leptons via displaced-vertex searches from ALP decays in D-meson production.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages · cited by 1 Pith paper

[1]

Achasov, X

M. Achasov, X. Ai, L. An, R. Aliberti, Q. An, X. Bai et al.,STCF conceptual design report (volume 1): Physics& detector,Frontiers of Physics19(2024) 14701

work page 2024
[2]

Z. Fang, Y. Pei, H. Shi, H. Peng and J. Liu,Study on background full simulation and implementation for Super Tau-Charm Facility, Journal of Instrumentation19(2024) P11014

work page 2024
[3]

H. Zhou, K. Sun, Z. Lu, H. Li, X. Ai, J. Zhang et al.,Global track finding based on the hough transform in the STCF detector,Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment1075 (2025) 170357

work page 2025
[4]

Höppner, S

C. Höppner, S. Neubert, B. Ketzer and S. Paul,A novel generic framework for track fitting in complex detector systems, Nuclear Inst& Methods in Physics Research A620 (2009) 518

work page 2009
[5]

Amrouche, L

S. Amrouche, L. Basara, P. Calafiura, D. Emeliyanov, V. Estrade, S. Farrell et al.,The tracking machine learning challenge: throughput phase, Computing and Software for Big Science7 (2023) 1

work page 2023
[6]

Tsaris, D

A. Tsaris, D. Anderson, J. Bendavid, P. Calafiura, G. Cerati, J. Esseiva et al.,The HEP.TrkX project: deep learning for particle tracking, inJournal of Physics: Conference Series, vol. 1085, p. 042023, IOP Publishing, 2018, DOI

work page 2018
[7]

Farrell, P

S. Farrell, P. Calafiura, M. Mudigonda, D. Anderson, J.-R. Vlimant, S. Zheng et al.,Novel deep learning methods for track reconstruction,arXiv preprint arXiv:1810.06111(2018)

arXiv 2018
[8]

Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang and P.S. Yu,A comprehensive survey on graph neural networks, IEEE transactions on neural networks and learning systems32(2020) 4. – 10 –

work page 2020

Show all 21 references

[9]

Caillou, C

S. Caillou, C. Collard, C. Rougier, J. Stark, H. Torres and A. Vallier,Novel fully-heterogeneous GNN designs for track reconstruction at the HL-LHC, inEPJ Web of Conferences, vol. 295, p. 09028, EDP Sciences, 2024, DOI

2024
[10]

Choma, D

N. Choma, D. Murnane, X. Ju, P. Calafiura, S. Conlon, S. Farrell et al.,Track seeding and labelling with embedded-space graph neural networks,arXiv preprint arXiv:2007.00149(2020)

2020 arXiv
[11]

M. Liu, W. Li, X. Huang, Y. Zhang, T. Lin and Y. Yuan,Simulation and reconstruction of particle trajectories in the CEPC drift chamber, inNuclear Science and Techniques, vol. 35, 2024, DOI

2024
[12]

Ablikim, Z

M. Ablikim, Z. An, J. Bai, N. Berger, J. Bian, X. Cai et al.,Design and construction of the BESIII detector, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment614 (2010) 345

2010
[13]

X. Jia, X. Qin, T. Li, X. Huang, X. Zhang, N. Yin et al.,BESIII track reconstruction algorithm based on machine learning, inEPJ Web of Conferences, vol. 295, p. 09006, EDP Sciences, 2024, DOI

2024
[14]

H. Zhou, Z. Fang, J. Liu and Y. Zhou,Design and simulation of a cylindrical𝜇RWELL inner tracker for the experiment at the Super Tau-Charm Facility, Journal of Instrumentation18(2023) P06015

2023
[15]

X. Ai, X. Huang, T. Li, B. Qi and X. Qin,Design and development of STCF offline software,Modern Physics Letters A(2024) 2440006

2024
[16]

Huang, T

W. Huang, T. Li, Q. Li, H. Li, D. Liu and X. Huang,Offline software framework for the super tau charm facility, inJournal of Physics: Conference Series, vol. 2438, p. 012054, IOP Publishing, 2023, DOI

2023
[17]

X. Ai, C. Allaire, N. Calace, A. Czirkos, M. Elsing, I. Ene et al.,A common tracking software project, Computing and Software for Big Science6 (2022) 8

2022
[18]

X. Ai, X. Huang and Y. Liu,Implementation of ACTS for STCF track reconstruction,Journal of Instrumentation 18(2023) P07026

2023
[19]

H. Li, H. Zhou, J. Zhang, X. Huang, J. Yang and X. Ai,Studies of track finding for long-lived particles at the super tau-charm facility,Nuclear Science and Techniques36(2025)

2025
[20]

Kingma and J

D.P. Kingma and J. Ba,Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[21]

Agostinelli, J

S. Agostinelli, J. Allison, K.a. Amako, J. Apostolakis, H. Araujo, P. Arce et al.,GEANT4—a simulation toolkit,Nuclear instruments and methods in physics research section A: Accelerators, Spectrometers, Detectors and Associated Equipment506 (2003) 250. – 11 –

2003

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Achasov, X

M. Achasov, X. Ai, L. An, R. Aliberti, Q. An, X. Bai et al.,STCF conceptual design report (volume 1): Physics& detector,Frontiers of Physics19(2024) 14701

work page 2024

[2] [2]

Z. Fang, Y. Pei, H. Shi, H. Peng and J. Liu,Study on background full simulation and implementation for Super Tau-Charm Facility, Journal of Instrumentation19(2024) P11014

work page 2024

[3] [3]

H. Zhou, K. Sun, Z. Lu, H. Li, X. Ai, J. Zhang et al.,Global track finding based on the hough transform in the STCF detector,Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment1075 (2025) 170357

work page 2025

[4] [4]

Höppner, S

C. Höppner, S. Neubert, B. Ketzer and S. Paul,A novel generic framework for track fitting in complex detector systems, Nuclear Inst& Methods in Physics Research A620 (2009) 518

work page 2009

[5] [5]

Amrouche, L

S. Amrouche, L. Basara, P. Calafiura, D. Emeliyanov, V. Estrade, S. Farrell et al.,The tracking machine learning challenge: throughput phase, Computing and Software for Big Science7 (2023) 1

work page 2023

[6] [6]

Tsaris, D

A. Tsaris, D. Anderson, J. Bendavid, P. Calafiura, G. Cerati, J. Esseiva et al.,The HEP.TrkX project: deep learning for particle tracking, inJournal of Physics: Conference Series, vol. 1085, p. 042023, IOP Publishing, 2018, DOI

work page 2018

[7] [7]

Farrell, P

S. Farrell, P. Calafiura, M. Mudigonda, D. Anderson, J.-R. Vlimant, S. Zheng et al.,Novel deep learning methods for track reconstruction,arXiv preprint arXiv:1810.06111(2018)

arXiv 2018

[8] [8]

Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang and P.S. Yu,A comprehensive survey on graph neural networks, IEEE transactions on neural networks and learning systems32(2020) 4. – 10 –

work page 2020