Pith. sign in

REVIEW 4 major objections 5 minor 12 references

TPCpp-10M: Simulated proton-proton collisions in a Time Projection Chamber for AI Foundation Models

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read TPCpp-10M provides 10 million simulated proton–proton collisions as TPC spacepoints in a common array format, with 70,000 labeled events for track finding, particle identification, and noise tagging.

desk verdict Useful public TPC simulation dataset, but its 'noise tagging' benchmark is mislabeled and needs a fix before serious use. read the letter →

arxiv 2509.05792 v1 pith:JFKUV6DL submitted 2025-09-06 physics.data-an

Shuhang Li , Yi Huang , David Park , Xihaier Luo , Haiwang Yu , Yeonju Go , Christopher Pinkenburg , Yuewei Lin
show 5 more authors
Shinjae Yoo Joseph Osborn Christof Roland Jin Huang Yihui Ren
This is my paper · ORCID
classification physics.data-an
keywords timeprojectionchamberfoundationmodelsself-supervisedlearningtrackfindingparticleidentificationnoisetaggingbenchmarkdatasetsimulatedcollisions
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

TPCpp-10M is a public dataset of 10 million simulated proton–proton collisions at 200 GeV, stored as time projection chamber (TPC) spacepoints in a common array format. Ten million unlabeled events are intended for self-supervised pretraining of machine-learning foundation models, while 70,000 labeled events carry ground-truth labels for track finding, particle identification, and noise tagging. The dataset is designed to give machine-learning researchers a low-barrier, reproducible testbed for TPC tracking, complementing public tracking datasets that come mostly from silicon detectors. The paper's central claim is that this resource is large and faithful enough to support scaling studies and transfer of foundation models to nuclear and particle physics.

What carries the argument

The central object is the TPC spacepoint: a 4-entry record (E, x, y, z) of an ionization signal's position and amplitude left by a charged particle in a gas-filled time projection chamber. The unlabeled dataset contains only these arrays; the labeled dataset aligns them with truth track IDs, noise tags, and particle identities. Because every downstream task—pretraining, track finding, PID, noise tagging—is expressed as operations on the same spacepoint arrays, the spacepoint is the common representational unit that connects raw simulation to foundation-model training. The evaluation machinery is double-majority matching, which declares a predicted track matched to a true particle only when h

What would settle it

Compare the spacepoint multiplicity, noise occupancy, and hit-position residuals in this dataset against real minimum-bias proton–proton data from the same TPC at 200 GeV; a large disagreement in any of these distributions would break the claim that the simulation is a faithful surrogate.

Watch

Extended reading notes

Core claim

The paper's central claim is that a 10-million-event simulated collision sample, stored as TPC spacepoints, can serve as the training and evaluation substrate for AI foundation models in charged-particle tracking. Each event is reduced to 4-column spacepoint records—ionization signal plus x, y, z position—plus, in the labeled subset, aligned track IDs, noise tags, and five-class particle-type labels. The authors assert the events are generated with a standard Monte Carlo event generator at 200 GeV, transported through a full detector simulation, digitized with channel-dependent gain and noise, and clustered into spacepoints, with the full chain reproducible from pinned software releases. For

Load-bearing premise

The load-bearing premise is that the simulated detector response—material, electronics, noise, and clustering—produces spacepoints that are a realistic stand-in for real TPC collisions; no comparison with real data is included.

Editorial extensions

If this is right

  • Machine-learning groups can pretrain track-finding models on 10 million events without running particle-physics software or learning physics-specific data formats.
  • The 70,000 labeled events give the community a common benchmark: track finding, PID, and noise tagging can be compared across methods using the recommended metrics.
  • Because the full simulation and reconstruction chain is pinned to specific releases, the dataset can be regenerated or extended consistently when detector settings change.
  • TPC-specific effects—long drift distances, diffusion, and up to roughly 100 truth tracks per event—become testable for models trained on silicon-tracker data.

Reading between the lines

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

  • An extension the paper leaves implicit: using only part of the 10M unlabeled events and measuring downstream task performance would directly probe scaling laws for self-supervised TPC representations.
  • The noise label is defined as spacepoints whose matched track has pT below 60 MeV/c; a benchmark built on this label may behave differently from one built on unmatched reconstructed fake spacepoints, so results should be read with this definition in mind.
  • A natural next experiment is testing whether models pretrained on these simulated spacepoints transfer to real TPC data; the release does not include a real-data comparison, so that transfer remains open.
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.

Referee Report

4 major / 5 minor

Summary. The paper introduces TPCpp-10M, a dataset of 10 million simulated proton-proton collisions at sqrt(s)=200 GeV for the sPHENIX TPC, generated with Pythia8 and Geant4 through the sPHENIX software stack. The dataset is released in NumPy format, with an unlabeled set for self-supervised pretraining and a labeled subset for track finding, particle identification, and noise tagging. The authors provide repository URLs, commit hashes, and an sPHENIX build release number for reproducibility, and they recommend evaluation metrics for each downstream task. The central claim is that this is a large, openly accessible, high-fidelity TPC dataset that lowers the barrier for machine learning researchers to work on particle tracking.

Significance. If the dataset is indeed of the claimed size and quality, it fills a real gap: most public tracking datasets are silicon-based, whereas TPC data have different pattern-recognition challenges (long drift, diffusion, many layers). The 10M-event scale is well suited for self-supervised foundation-model training, and the NumPy format plus example scripts make it accessible outside the particle-physics community. The reproducibility measures are a strong point: commit hashes plus the CVMFS release (ana.435) give a concrete, auditable generation pipeline. The three labeled tasks are sensible in principle, and the double-majority tracking metric is a standard choice. The main value is the resource itself; however, several specific label and consistency issues in the labeled subset must be addressed before the benchmark claims are reliable.

major comments (4)
  1. [Data Description / labeled folder structure] The labeled-set size is internally inconsistent. The abstract and Value of the Data state 70,000 labeled events, but the Data Description (and Figure 3) list 7 train shards of 10k events each (70k), plus a 13k validation file and a 7k test file, totaling 90k. This discrepancy affects every downstream benchmark comparison, because readers cannot know the actual train/validation/test sizes or whether the abstract is undercounting. Please state exactly how many events are in each split and correct the abstract/value statement accordingly.
  2. [Data Description, noise_tags definition; Figure 5(b)] The noise-tagging task is mislabeled. noise_tags is defined as 1 iff the matched track's pT < 60 MeV/c and 0 otherwise, and the task description says it flags noise. This is a kinematic cut on truth-matched tracks, not a detector-noise flag. Real TPC noise (electronics pickup, unmatched clusters, random combinations) has no truth track and is not captured by this rule. The stated rationale that such tracks are 'unlikely to reach the TPC active volume' is also quantitatively odd: a 60 MeV/c primary in the 1.4 T field has gyroradius ~14 cm, well below the TPC inner radius of ~32 cm, so it cannot reach the active volume at all; the tagged particles must be secondaries or conversions. One of the three downstream tasks is therefore not what it claims to be. Either rename the task to 'low-pT track tagging' or provide genuine unmatched/noise hits as the positive class.
  3. [Limitations / Experimental Design] There is no validation against real sPHENIX collision data. The manuscript claims 'realistic detector conditions' and 'high-fidelity resource,' but no comparison is shown of spacepoint multiplicities, hit resolutions, dE/dx distributions, or tracking performance between simulation and real TPC data. The Limitations section only says 'Residual differences from real data may exist.' For a dataset whose primary purpose is to stand in for real data in foundation-model training, this is a load-bearing gap. Please add at least basic quantitative cross-checks (e.g., spacepoint distributions, cluster shapes, or tracking efficiency against available sPHENIX p+p reference data) or explicitly qualify the dataset as simulation-only with no fidelity claim.
  4. [Data preprocessing / labeled subset applicability] The preprocessing section states that tracks with fewer than five spacepoints are removed and events with fewer than 20 remaining spacepoints are dropped. It is not stated whether these cuts are applied to the unlabeled 10M set, the labeled 90k set, or both. If the labeled noise-tagging task is meant to include detector noise, removing short tracks and requiring every spacepoint to belong to a truth track would already eliminate the very objects that should be labeled as noise. Clarify the filter scope and report how many events were rejected at each step.
minor comments (5)
  1. [Abstract / Data Description] The mismatch between 70k and 90k should be fixed everywhere it appears; see major comment.
  2. [Data Description, noise_tags wording] Even after renaming, the phrase 'unlikely to reach the TPC active volume' should be replaced with a precise statement about production radius and pT thresholds, since a primary track at 60 MeV/c cannot reach the TPC at all.
  3. [Experimental Design] Minor typo: 'detector and1.4 Tmagnetic field' is missing a space after 'and'.
  4. [Figure 4 caption] The caption repeats 'from fewer than 10 to nearly 100 truth tracks per event' twice; the second occurrence should be removed or reworded.
  5. [Evaluation Metrics] The double-majority matching section would benefit from a one-sentence note on how to treat candidates/particles with zero hits in the denominator, though this is standard practice and not blocking.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; dataset resource with no derivation chain to reduce.

full rationale

TPCpp-10M is a data resource paper rather than a derivation. It reports a generated dataset (Pythia8 + Geant4 + sPHENIX digitization/reconstruction), labels derived from truth associations, and recommended metrics. There is no equation in which an output is defined as an input, no fitted parameter renamed as a prediction, and no uniqueness theorem imported from the authors' prior work. The references to the sPHENIX software stack ([8]–[11]) cite public repositories with specific commit hashes and a CVMFS release (ana.435), so they serve as reproducibility artifacts rather than as self-citations carrying a load-bearing argument. The only potentially problematic passage is the 'noise tagging' label definition: 'noise_tags … equal to 1 if the matched track’s transverse momentum pT < 60 MeV/c and 0 otherwise. This flags spacepoints unlikely to reach the TPC active volume.' This is a kinematic cut rather than a genuine detector-noise flag, and the 'Data preprocessing' section removes short tracks that 'almost always originate from detector noise,' which is internally inconsistent. However, that is a labeling-validity/correctness concern, not circularity: the label is generated from stated truth criteria and does not reduce a predicted quantity to an input. The Limitations passage ('Residual differences from real data may exist') appropriately concedes the absence of data/simulation comparison. The recommended metrics are self-defined but clearly stated, which is standard for benchmark resources. Overall, no circular step is present, so the score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim of a useful dataset depends on simulation fidelity, the appropriateness of hand-chosen cuts (pT threshold, track length, event occupancy), and the correctness of truth associations. None of these are externally validated in the paper, and the free parameters affect the balance of all three benchmark tasks.

free parameters (3)
  • noise_tag_pT_cutoff = 60 MeV/c
    The definition of noise tags uses pT < 60 MeV/c as a threshold. This is a hand-chosen cutoff that determines the noise class composition and the difficulty of the noise-tagging benchmark.
  • min_spacepoints_per_track = 5
    Tracks with fewer than five spacepoints are removed. This cut is described as removing short segments that 'almost always originate from detector noise', but the threshold is arbitrary and changes the event composition.
  • min_spacepoints_per_event = 20
    Events with fewer than 20 spacepoints after track removal are rejected, reducing the sample from raw minimum-bias events. The number of rejected events is not reported.
assumptions (3)
  • domain assumption The sPHENIX Geant4 simulation with FTFP_BERT_HP physics list accurately models the detector response and ionization signals.
    Invoked throughout the Experimental Design section, especially the 'TPC response, digitization, and reconstruction' paragraph, with no validation against real collision data.
  • domain assumption Pythia8.307 with the Detroit tune provides a faithful model of p+p collisions at 200 GeV for minimum-bias events.
    Used as the event generator and cited from references [1] and [2]; the authors assume its output is adequate for ML training data.
  • domain assumption Spacepoint-to-truth associations produced by the simulation and clustering algorithms are correct and unambiguous.
    The labeled tasks rely on these associations being accurate; the paper does not report any validation of clustering truth-matching efficiency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TPCpp-10M: Simulated proton-proton collisions in a Time Projection Chamber for AI Foundation Models." pith.science (2026). https://pith.science/paper/JFKUV6DL

@misc{pith2026250905792,
  author       = {Pith},
  title        = {Pith review of: TPCpp-10M: Simulated proton-proton collisions in a Time Projection Chamber for AI Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFKUV6DL}},
  note         = {Machine review of arXiv:2509.05792}
}
read the original abstract

Scientific foundation models hold great promise for advancing nuclear and particle physics by improving analysis precision and accelerating discovery. Yet, progress in this field is often limited by the lack of openly available large scale datasets, as well as standardized evaluation tasks and metrics. Furthermore, the specialized knowledge and software typically required to process particle physics data pose significant barriers to interdisciplinary collaboration with the broader machine learning community. This work introduces a large, openly accessible dataset of 10 million simulated proton-proton collisions, designed to support self-supervised training of foundation models. To facilitate ease of use, the dataset is provided in a common NumPy format. In addition, it includes 70,000 labeled examples spanning three well defined downstream tasks: track finding, particle identification, and noise tagging, to enable systematic evaluation of the foundation model's adaptability. The simulated data are generated using the Pythia Monte Carlo event generator at a center of mass energy of sqrt(s) = 200 GeV and processed with Geant4 to include realistic detector conditions and signal emulation in the sPHENIX Time Projection Chamber at the Relativistic Heavy Ion Collider, located at Brookhaven National Laboratory. This dataset resource establishes a common ground for interdisciplinary research, enabling machine learning scientists and physicists alike to explore scaling behaviors, assess transferability, and accelerate progress toward foundation models in nuclear and high energy physics. The complete simulation and reconstruction chain is reproducible with the sPHENIX software stack. All data and code locations are provided under Data Accessibility.

Figures

Figures reproduced from arXiv: 2509.05792 by the authors.

Figure 1
Figure 1. RHIC and sPHENIX experiment (context figure). Background In nuclear and particle physics experiments, tracking detectors are used to measure and reconstruct the trajectories of charged particles. These measurements are crucial for determining momenta, identifying particle species, and studying underlying physics processes. The TPC provides three-dimensional (3D) spacepoints that reconstruct the trajectories of parti… view at source ↗
Figure 2
Figure 2. TPC schematic and collision event. Panel (a): The TPC schematic, where proton beams [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Folder structure of the TPCpp-10M dataset. Each event includes reconstructed TPC spacepoints, as well as (for the labeled set) truth associations and per-track kinematics. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Panel (a): Distribution of spacepoints per event. Panel (b): distribution of truth tracks per [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Downstream tasks ground truth labels. Track finding: assign every reconstructed spacepoint to a track identity using the pro￾vided truth associations. The label for each spacepoint is the integer track_id of its matched truth track. Note: Evaluation may be restricted t…
Figure 6
Figure 6. Figure 6: Box plots of per-event ground truth class ratios in the test subset of the labeled dataset, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 8 canonical work pages

  1. [1]

    Sjöstrand, S

    T. Sjöstrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten, S. Mrenna, S.Prestel, C.O.Rasmussen, P.Z.Skands, AnintroductiontoPYTHIA8.2, Comput. Phys.Commun.191(2015)159–177.arXiv:1410.3012,doi:10.1016/j.cpc.2015. 01.024

    arXiv 2015

  2. [2]

    M. R. Aguilar, Z. Chang, R. K. Elayavalli, R. Fatemi, Y. He, Y. Ji, D. Kalinkin, M. Kelsey, I. Mooney, V. Verkest, pythia8 underlying event tune for RHIC energies, Phys. Rev. D 105 (1) (2022) 016011.arXiv:2110.09447,doi:10.1103/PhysRevD. 105.016011

    arXiv 2022

  3. [3]

    Agostinelli, et al., GEANT4 - A Simulation Toolkit, Nucl

    S. Agostinelli, et al., GEANT4 - A Simulation Toolkit, Nucl. Instrum. Meth. A 506 (2003) 250–303.doi:10.1016/S0168-9002(03)01368-8

    doi:10.1016/s0168-9002(03)01368-8 2003

  4. [4]

    Klest, Overview and design of the sphenix tpc, Journal of Physics: Conference Series 1498 (1) (2020) 012025.doi:10.1088/1742-6596/1498/1/012025

    H. Klest, Overview and design of the sphenix tpc, Journal of Physics: Conference Series 1498 (1) (2020) 012025.doi:10.1088/1742-6596/1498/1/012025. URLhttps://dx.doi.org/10.1088/1742-6596/1498/1/012025

    work page doi:10.1088/1742-6596/1498/1/012025 2020

  5. [5]

    Rev.D 110(3) (2024)030001

    Particle DataGroup, Review ofparticle physics, Phys. Rev.D 110(3) (2024)030001. doi:10.1103/PhysRevD.110.030001

    doi:10.1103/physrevd.110.030001 2024

  6. [6]

    X. Ju, D. Murnane, P. Calafiura, N. Choma, S. Conlon, S. Farrell, Y. Xu, M. Spirop- ulu, J.-R. Vlimant, A. Aurisano, et al., Performance of a geometric deep learning pipeline for hl-lhc particle tracking, The European Physical Journal C 81 (2021) 1–14

    work page 2021

  7. [7]

    Garcia, M

    D. Garcia, M. Selvaggi, B. Francois, Geometric graph neural network based track finding, openreview (2025)

    work page 2025

  8. [8]

    sPHENIX Collaboration, coresoftware: Core simulation and reconstruction software for the sphenix experiment, tagged atb849eba5c2cf8ada510d036aa9b9499cb31f0513,https: //github.com/sPHENIX-Collaboration/coresoftware/tree/ b849eba5c2cf8ada510d036aa9b9499cb31f0513, accessed August 3, 2025; used in the official sPHENIX build, including downstream emulation and...

    work page 2025

Show all 12 references
  1. [9]

    sPHENIX Collaboration, acts: A common tracking software toolkit for sphenix, tagged at33fc284f238a24405bcd6c2de3260f370d6f8403, https://github.com/sPHENIX-Collaboration/acts/tree/ 33fc284f238a24405bcd6c2de3260f370d6f8403, accessed August 3, 2025; part of the official sPHENIX b...

    2025

  2. [10]

    sPHENIX Collaboration, macros: Analysis and utility macros for sphenix, tagged at661f781db23352a3fa72055ce7dbf1a0ee1c2167, https://github.com/sPHENIX-Collaboration/macros/tree/ 661f781db23352a3fa72055ce7dbf1a0ee1c2167, accessed August 3, 2025; part of the official sPHENIX buil...

    2025

  3. [11]

    sPHENIX Collaboration, calibrations: Calibration code for sphenix, tagged ata3e66e69635514813ee3e20bf18b2bd59787b503, https://github.com/sPHENIX-Collaboration/calibrations/tree/ a3e66e69635514813ee3e20bf18b2bd59787b503, accessed August 3, 2025; part of the official sPHENIX bui...

    2025

  4. [12]

    Blomer, P

    J. Blomer, P. Buncic, T. Fuhrmann, CernVM-FS: delivering scientific software to globally distributed computing resources, in: International Conference for High Performance Computing, Networking, Storage and Analysis, 2011.doi:10.1145/ 2110217.2110225. 11

    2011

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.