Pith. sign in

REVIEW 3 major objections 4 minor 66 references

A GPT-like transformer can generate silicon tracker hits with quality close to full simulation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 13:22 UTC pith:EAUVYCPA

load-bearing objection First real shot at ML-based silicon tracker simulation, with honest limitations inside the paper but an abstract that oversells 'comparable' — and the 3-hit context window is a plausible culprit that goes untested. the 3 major comments →

arxiv 2512.24254 v2 pith:EAUVYCPA submitted 2025-12-30 physics.ins-det hep-ex

GPT-like transformer model for silicon tracking detector simulation

Tadej Novak , Borut Paul Ker\v{s}evan This is my paper
classification physics.ins-det hep-ex PACS 29.40.Gx
keywords silicon tracking detectorgenerative machine learningtransformerGPTdetector simulationtokenizationOpen Data Detectortrack reconstruction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper establishes that a decoder-only transformer, of the kind used for text generation, can generate the sequence of silicon tracker hits left by a particle as it crosses a detector, and that these synthetic hits reconstruct into tracks of nearly the same quality as hits from full simulation. Using the Open Data Detector and single muons, the model reaches 96.3% track fitting efficiency versus 98.1% for the rounded reference simulation and 99.9% for the unrounded one. The significance is practical: detector response simulation is one of the largest computing costs for the HL-LHC era, and fast generative models have so far been applied almost exclusively to calorimeters. The paper also shows the approach's limits—rare processes such as pion decays inside the tracker and hard electron bremsstrahlung are underproduced—so the central claim is 'comparable, not identical, to full simulation.'

Core claim

The central discovery is that a GPT-like autoregressive transformer, trained only on tokenized hit features, can act as a fully generative model of a silicon tracker: starting from a virtual start hit (beamspot position and initial momentum), it predicts each subsequent hit feature one token at a time, sampling from a constrained token set, until an end token is emitted. Evaluated on the Open Data Detector, single-muon tracks reconstructed from these generated hits have a technical track fitting efficiency of 96.3%, compared with 98.1% for the rounded Geant4 reference and 99.9% for full Geant4; hit-level distributions agree to a few percent. The authors argue this makes the transformer-based

What carries the argument

The load-bearing mechanism is a decoder-only transformer with masked self-attention, operating on a flat token sequence: every hit is flattened into seven feature values (particle ID, geometry ID, two local coordinates, three momentum components), each rounded and mapped to a shared token vocabulary, with offsets keeping discrete features unique. A virtual start hit and end token delimit the sequence. To keep inference tractable, training and generation use sliding windows of at most four hits, with an added hit-index feature; the model therefore predicts the next hit from at most three previous hits, an assumption justified by the claim that relevant correlations are local, dominated by tra

Load-bearing premise

The model assumes that three previous hits are enough context to predict the next hit, so any physics that depends on a longer track history, such as accumulated multiple scattering or a decay, can be missed.

What would settle it

Take simulated tracks with a deliberately large scattering at an early layer and compare the transformer-generated positions at later layers to full simulation; if the three-hit window is too short, the later hits will show systematic deviations from the true multiple-scattering history. Alternatively, measure whether curvature inferred from early generated hits and from late generated hits agrees; disagreement would show the short context breaks global consistency.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Fast simulation can be extended from calorimeters to silicon trackers, directly attacking the dominant computing cost of HL-LHC detector simulation.
  • Because the model generates the whole hit sequence autoregressively, correlations between hits are generated rather than imposed, so track-level quantities like curvature and scattering emerge from the sequence.
  • The tokenization precision is a controllable knob: rounding hits to two decimal places costs about one percentage point of tracking efficiency, so finer or detector-aware tokenization should recover most of the gap.
  • Larger models help: doubling the transformer dimension raised muon fitting efficiency from 94.9% to 96.3%, implying further scaling may close the remaining gap.
  • Rare processes (pion decay in the tracker, hard electron bremsstrahlung) are underproduced, so production use will require treating these as conditional or weighted cases rather than relying on the base model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the three-hit context window is truly sufficient, the same architecture could generate tracks in a streaming fashion, layer by layer, with bounded memory; a direct test would be to compare generated hits on tracks with artificially long autocorrelations (e.g., a large early scatter) against full simulation.
  • The flat-token representation naturally extends to whole events with secondary particles, but that will likely require hierarchical sequence models or explicit branching tokens, since a tree of secondary tracks is not a single linear sequence.
  • The observed drop in phi modeling at full 2π coverage hints that global geometry is the next bottleneck; conditioning on detector-region embeddings or a global position latent could be a testable fix.
  • Because the method learns correlations from data, it may transfer to other tracking geometries with only retraining, provided the token vocabulary and allowed-token masks are rebuilt—the Open Data Detector result is existence proof, not a guarantee.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents a GPT-like decoder-only transformer that generates silicon tracker hits as token sequences. It is trained on Geant4-simulated single muon, electron, and pion samples in the Open Data Detector, using a sliding window of four hits (three previous hits of context at inference). Hit features are tokenized after rounding continuous values to two decimals. The generated hits are evaluated both at hit level and through ACTS track reconstruction. For muons, hit-level distributions are close to rounded Geant4; the best model reaches 99.7% seeding and 96.3% fitting efficiency, versus 99.9%/98.1% for rounded Geant4 and 99.9%/99.9% for full Geant4. Electrons show momentum bias and pions rarely decay in the generated samples. GPU inference is considerably faster than CPU-based Geant4.

Significance. This is an interesting first demonstration of generative sequence modeling for silicon tracking detectors, with openly released code and data, and validation through a standard reconstruction chain rather than only distribution-level metrics. The tokenized track representation and the use of ACTS as an external evaluator are strengths, and the paper is transparent about several limitations. However, the central claim of 'comparable with the full simulation' is not supported by the reported numbers: the best transformer is 1.8% below rounded Geant4 and 3.6% below full Geant4 in fitting efficiency, and the electron/pion results are clearly worse. The load-bearing assumption of a three-hit context window is asserted but never ablated. As a proof-of-principle the paper is valuable; as a claim of comparable tracking performance it needs substantial qualification and additional experiments.

major comments (3)
  1. [Abstract; §5.1, Table 3] The abstract states that the tracking performance 'is comparable with the full simulation.' This is not supported by Table 3: for the best muon model (35.0 M), fitting efficiency is 96.3% versus 98.1% for rounded Geant4 and 99.9% for full Geant4; seeding is 99.7% versus 99.9%. The gap is larger than the rounding effect alone (1.8%). For electrons and pions, §5.2 reports biased momentum and missing decays. Please qualify the claim (e.g., 'comparable to rounded Geant4 for muons, with known limitations for electrons and pions').
  2. [§4 (sliding-window attention)] The inference uses at most three previous hits of context. This is the key generative assumption, justified only by the assertion that hit correlations drop with distance. No ablation or measured correlation length is given. The observed efficiency loss could stem from truncated context, since track parameters are global and multiple scattering creates memory beyond three hits. Please add an experiment varying window size (e.g., full-sequence attention, 4/6/8 hits) and report both validation loss and tracking efficiency. This is needed to support the claim that the chosen architecture is adequate.
  3. [§5.2] The conclusion that 'overall track quality is comparable' for electrons and pions is not backed by a track-level efficiency table analogous to Table 3. Only hit-level distributions and qualitative statements are provided. Without numbers, the reader cannot judge whether these samples meet the stated goal. Either add seeding/fitting efficiencies for electrons and pions or explicitly state that track-level validation was performed only for muons and restrict the abstract accordingly.
minor comments (4)
  1. [Figures 4, 5, 7, 10] Several figure captions contain garbled text ('par icles', '/uni03D5', 'T able', 'muons' with stray characters). Please replace with clean, correctly rendered versions.
  2. [§5.1] The text says the smallest benchmark model reaches comparable tracking performance to rounded Geant4, but Table 3 only lists the 11.2 M and 35.0 M models. Report the missing efficiency values for the small benchmark model or clarify that the comparison is across different phase-space selections.
  3. [§5.3, Table 5] The claim that bf16 precision gives 'no physics performance change' is not supported by Table 5, which lists only training/inference times. Add a pointer to the physics comparison or soften the claim.
  4. [Table 4] The Geant4 reference rows should be annotated with the same units as the transformer rows ('per 10,000 simulated tracks') to make the comparison unambiguous.

Circularity Check

0 steps flagged

No significant circularity: the transformer is trained on held-out Geant4 data and its tracking performance is benchmarked externally with ACTS; no target tracking metric enters training or model selection.

full rationale

The central claim is that GPT-like transformers can generate silicon tracker hits whose reconstructed track quality is close to Geant4. The derivation chain is: Geant4/ODD simulation -> tokenization/rounding -> autoregressive next-token training by cross-entropy -> inference from the starting virtual hit -> comparison of hit distributions and ACTS seeding/fitting efficiencies against Geant4 on a held-out test sample. No circularity pattern is present. Model selection is explicitly by validation loss ('The trained model yielding the lowest validation loss is taken'), and the validation loop is separate from the tracking metrics reported in Table 3. The tracking efficiency is not fitted: it is measured only after generation, using the external ACTS reference configuration, against both full and rounded Geant4 on test events not seen in training ('The inference is performed on the test sample to ensure that only events never seen by the neural network are used for validation'). The paper also transparently states that its results are compared with rounded Geant4, the same quantized representation used for training, while still reporting the unrounded reference. The only self-references are code-availability citations (SiliconAI and the validator); no load-bearing mathematical claim rests on a self-citation or on a uniqueness theorem imported from the authors. The sliding-window/three-hit context and the failure to model rare processes (pion decay, electron bremsstrahlung tails) are genuine limitations and potential correctness risks, but they are not cases where a prediction reduces by construction to its input. In fact, the observed degraded performance for electrons/pions and the quantisation-induced efficiency drop are evidence that the evaluation is not automatically reproducing a fitted target.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physics entities are introduced. The free parameters are modeling choices that directly affect fidelity. The axioms are domain assumptions about the simulation chain and the adequacy of the simplified representation.

free parameters (3)
  • Rounding precision for continuous features = 2 decimal places
    Continuous features are rounded to two decimals to limit token space; this degrades tracking efficiency by ~1% (Section 5.1).
  • Sliding window size = 4 hits (context of 3 previous hits at inference)
    Chosen to reduce sequence length under the assumption that local context suffices; affects correlation capture.
  • Maximum sequence length = 40
    Set to be divisible by the number of attention heads; imposes an upper limit on generated hit count.
axioms (5)
  • domain assumption Geant4 simulation provides the ground-truth detector response.
    The entire training and validation uses Geant4 output as reference; if Geant4 is inaccurate, the model inherits those inaccuracies.
  • domain assumption Technical seeding and fitting efficiencies are valid proxies for physics performance.
    The paper uses technical efficiencies to isolate algorithmic performance, assuming they sufficiently reflect physics performance.
  • domain assumption Secondary particles can be discarded without losing essential features of the primary track.
    This simplification enables the sequence formulation but fails for processes like pion decay, as the paper acknowledges.
  • domain assumption A fixed deterministic ordering of hits and features is sufficient for the generative task.
    The model relies on a flat feature sequence with a fixed order; if detector geometry requires permutation-invariant treatment, this would break.
  • domain assumption A short context window captures all relevant inter-hit correlations.
    Justified by locality of track curvature; not validated for long-range effects such as multiple scattering or material-dependent correlations.

pith-pipeline@v1.3.0-alltime-deepseek · 15192 in / 8396 out tokens · 87117 ms · 2026-08-03T13:22:49.308109+00:00 · methodology

0 comments
read the original abstract

Simulating physics processes and detector responses is essential in high energy physics and represents significant computing costs. Generative machine learning has been demonstrated to be potentially powerful in accelerating simulations, outperforming traditional fast simulation methods. The efforts have focused primarily on calorimeters. This work presents the very first studies on using neural networks for silicon tracking detectors simulation. The GPT-like transformer architecture is determined to be optimal for this task and applied in a fully generative way, ensuring full correlations between individual hits. Taking parallels from text generation, hits are represented as a flat sequence of feature values. The resulting tracking performance, evaluated on the Open Data Detector, is comparable with the full simulation.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 3 canonical work pages

  1. [1]

    Technical report, CERN, Geneva (2022)

    Boehnlein, A., Biscarat, C., Bressan, A., Britton, D., Bolton, R., Gaede, F., Grandi, C., Hernandez, F., Kuhr, T., Merino, G., Simon, F., Watts, G.: HL-LHC Software and Computing Review Panel, 2nd Report. Technical report, CERN, Geneva (2022). https://cds.cern.ch/record/2803119

    arXiv 2022

  2. [2]

    : A Roadmap for HEP Software and Computing R&D for the 2020s

    Albrecht, J., et al. : A Roadmap for HEP Software and Computing R&D for the 2020s. Comput. Softw. Big Sci. 3(1), 7 (2019) https://doi.org/10.1007/s41781-018-0018-8 arXiv:1712.06982 [physics.comp-ph]

    Pith/arXiv arXiv 2019

  3. [3]

    ATLAS Collaboration: The ATLAS Simulation Infrastructure. Eur. Phys. J. C 70, 823–874 (2010) https://doi.org/10.1140/epjc/s10052-010-1429-9 arXiv:1005.4568 [physics.ins-det]

    Pith/arXiv arXiv 2010

  4. [4]

    Apostolakis, J., et al.: HEP Software Foun- dation Community White Paper Work- ing Group - Detector Simulation (2018) arXiv:1803.04165 [physics.comp-ph]

    Pith/arXiv arXiv 2018

  5. [5]

    : GEANT4 - A Simulation Toolkit

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

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

  6. [6]

    Paganini, M., Oliveira, L., Nachman, B.: Accelerating Science with Generative Adver- sarial Networks: An Application to 3D Particle Showers in Multilayer Calorime- ters. Phys. Rev. Lett. 120(4), 042003 (2018) https://doi.org/10.1103/PhysRevLett.120.042003 arXiv:1705.02355 [hep-ex]

    Pith/arXiv arXiv 2018

  7. [7]

    Paganini, M., Oliveira, L., Nachman, B.: CaloGAN : Simulating 3D high energy par- ticle showers in multilayer electromagnetic calorimeters with generative adversarial net- works. Phys. Rev. D 97(1), 014021 (2018) https://doi.org/10.1103/PhysRevD.97.014021 arXiv:1712.10321 [hep-ex]

    Pith/arXiv arXiv 2018

  8. [8]

    Erdmann, M., Geiger, L., Glombitza, J., Schmidt, D.: Generating and refining particle detector simulations using the Wasser- stein distance in adversarial networks. Comput. Softw. Big Sci. 2(1), 4 (2018) https://doi.org/10.1007/s41781-018-0008-x arXiv:1802.03325 [astro-ph.IM]

    Pith/arXiv arXiv 2018

  9. [9]

    Erdmann, M., Glombitza, J., Quast, T.: Precise simulation of electromagnetic calorimeter showers using a Wasser- stein Generative Adversarial Network. Comput. Softw. Big Sci. 3(1), 4 (2019) 11 https://doi.org/10.1007/s41781-018-0019-7 arXiv:1807.01954 [physics.ins-det]

    Pith/arXiv arXiv 2019

  10. [10]

    Musella, P., Pandolfi, F.: Fast and Accu- rate Simulation of Particle Detectors Using Generative Adversarial Networks. Comput. Softw. Big Sci. 2(1), 8 (2018) https://doi.org/10.1007/s41781-018-0015-y arXiv:1805.00850 [hep-ex]

    Pith/arXiv arXiv 2018

  11. [11]

    : Calorimetry with deep learning: particle simulation and reconstruction for collider physics

    Belayneh, D., et al. : Calorimetry with deep learning: particle simulation and reconstruction for collider physics. Eur. Phys. J. C 80(7), 688 (2020) https://doi.org/10.1140/epjc/s10052-020-8251-9 arXiv:1912.06794 [physics.ins-det]

    Pith/arXiv arXiv 2020

  12. [12]

    SciPost Phys

    Butter, A., Diefenbacher, S., Kasieczka, G., Nachman, B., Plehn, T.: GANplifying event samples. SciPost Phys. 10(6), 139 (2021) https://doi.org/10.21468/SciPostPhys.10.6.139 arXiv:2008.06545 [hep-ph]

    Pith/arXiv arXiv 2021

  13. [13]

    ATLAS Collaboration: AtlFast3: The Next Generation of Fast Simulation in ATLAS. Comput. Softw. Big Sci. 6(1), 7 (2022) https://doi.org/10.1007/s41781-021-00079-7 arXiv:2109.02551 [hep-ex]

    Pith/arXiv arXiv 2022

  14. [14]

    Com- put

    ATLAS Collaboration: Deep Gen- erative Models for Fast Photon Shower Simulation in ATLAS. Com- put. Softw. Big Sci. 8(1), 7 (2024) https://doi.org/10.1007/s41781-023-00106-9 arXiv:2210.06204 [hep-ex]

    Pith/arXiv arXiv 2024

  15. [15]

    Nature Commun

    Hashemi, B., Hartmann, N., Sharifzadeh, S., Kahn, J., Kuhr, T.: Ultra-high-granularity detector simulation with intra-event aware generative adversarial network and self-supervised relational reasoning. Nature Commun. 15(1), 4916 (2024) https://doi.org/10.1038/s41467-024-49104-4 arXiv:2303.08046 [physics.ins-det]. [Erratum: Nature Commun. 115, 5825 (2024)]

    Pith/arXiv arXiv 2024

  16. [16]

    Faucci Giannelli, M., Zhang, R.: CaloShow- erGAN, a generative adversarial network model for fast calorimeter shower simula- tion. Eur. Phys. J. Plus 139(7), 597 (2024) https://doi.org/10.1140/epjp/s13360-024-05397-4 arXiv:2309.06515 [physics.ins-det]

    Pith/arXiv arXiv 2024

  17. [17]

    PTEP 2024(8), 083–01 (2024) https://doi.org/10.1093/ptep/ptae106 arXiv:2401.02248 [hep-ex]

    Simsek, E., Isildak, B., Dogru, A., Aydo- gan, R., Bayrak, A.B., Ertekin, S.: CALPAGAN: Calorimetry for Particles Using Generative Adversarial Net- works. PTEP 2024(8), 083–01 (2024) https://doi.org/10.1093/ptep/ptae106 arXiv:2401.02248 [hep-ex]

    Pith/arXiv arXiv 2024

  18. [18]

    Com- put

    Buhmann, E., Diefenbacher, S., Eren, E., Gaede, F., Kasieczka, G., Korol, A., Kr¨ uger, K.: Getting High: High Fidelity Simulation of High Granularity Calorimeters with High Speed. Com- put. Softw. Big Sci. 5(1), 13 (2021) https://doi.org/10.1007/s41781-021-00056-0 arXiv:2005.05334 [physics.ins-det]

    Pith/arXiv arXiv 2021

  19. [19]

    Buhmann, E., Diefenbacher, S., Hund- hausen, D., Kasieczka, G., Korcari, W., Eren, E., Gaede, F., Kr¨ uger, K., McKeown, P., Rustige, L.: Hadrons, better, faster, stronger. Mach. Learn. Sci. Tech. 3(2), 025014 (2022) https://doi.org/10.1088/2632-2153/ac7848 arXiv:2112.09709 [physics.ins-det]

    Pith/arXiv arXiv 2022

  20. [20]

    In: 36th Conference on Neural Information Processing Systems: Workshop on Machine Learning and the Physical Sciences (2022)

    Cresswell, J.C., Ross, B.L., Loaiza-Ganem, G., Reyes-Gonzalez, H., Letizia, M., Caterini, A.L.: CaloMan: Fast generation of calorime- ter showers with density estimation on learned manifolds. In: 36th Conference on Neural Information Processing Systems: Workshop on Machine Learning and the Physical Sciences (2022)

    2022

  21. [21]

    JINST 17(09), 09028 (2022) https://doi.org/10.1088/1748-0221/17/09/P09028 arXiv:2202.07352 [hep-ph]

    Bieringer, S., Butter, A., Diefenbacher, S., Eren, E., Gaede, F., Hundhausen, D., Kasieczka, G., Nachman, B., Plehn, T., Trabs, M.: Calomplification — the power of generative calorimeter models. JINST 17(09), 09028 (2022) https://doi.org/10.1088/1748-0221/17/09/P09028 arXiv:2202.07352 [hep-ph]

    Pith/arXiv arXiv 2022

  22. [22]

    Diefenbacher, S., Eren, E., Gaede, F., Kasieczka, G., Korol, A., Kr¨ uger, K., McKeown, P., Rustige, L.: New angles on fast calorimeter shower simulation. Mach. Learn. Sci. Tech. 4(3), 035044 (2023) https://doi.org/10.1088/2632-2153/acefa9 arXiv:2303.18150 [physics.ins-det]

    Pith/arXiv arXiv 2023

  23. [23]

    Hoque, S., Jia, H., Abhishek, A., Fadaie, 12 M., Toledo-Mar´ ın, J.Q., Vale, T., Melko, R.G., Swiatlowski, M., Fedorko, W.T.: CaloQVAE: Simulating high-energy particle- calorimeter interactions using hybrid quantum-classical generative models. Eur. Phys. J. C 84(12), 1244 (2024) https://doi.org/10.1140/epjc/s10052-024-13576-x arXiv:2312.03179 [hep-ex]

    arXiv 2024

  24. [24]

    Liu, Q., Shimmin, C., Liu, X., Shliz- erman, E., Li, S., Hsu, S.-C.: Calo- VQ: Vector-Quantized Two-Stage Genera- tive Model in Calorimeter Simulation (2024) arXiv:2405.06605 [physics.ins-det]

    Pith/arXiv arXiv 2024

  25. [25]

    Smith, D., Ghosh, A., Liu, J., Baldi, P., Whiteson, D.: Fast multi-geometry calorimeter simulation with conditional self- attention variational autoencoders (2024) arXiv:2411.05996 [hep-ex]

    Pith/arXiv arXiv 2024

  26. [26]

    Krause, C., Shih, D.: Fast and accurate simulations of calorime- ter showers with normalizing flows. Phys. Rev. D 107(11), 113003 (2023) https://doi.org/10.1103/PhysRevD.107.113003 arXiv:2106.05285 [physics.ins-det]

    Pith/arXiv arXiv 2023

  27. [27]

    Krause, C., Shih, D.: Accelerating accurate simulations of calorime- ter showers with normalizing flows and probability density distillation. Phys. Rev. D 107(11), 113004 (2023) https://doi.org/10.1103/PhysRevD.107.113004 arXiv:2110.11377 [physics.ins-det]

    Pith/arXiv arXiv 2023

  28. [28]

    Krause, C., Pang, I., Shih, D.: CaloFlow for CaloChallenge dataset

  29. [29]

    16(5), 126 (2024) https://doi.org/10.21468/SciPostPhys.16.5.126 arXiv:2210.14245 [physics.ins-det]

    SciPost Phys. 16(5), 126 (2024) https://doi.org/10.21468/SciPostPhys.16.5.126 arXiv:2210.14245 [physics.ins-det]

    Pith/arXiv arXiv 2024

  30. [30]

    JINST 18(10), 10017 (2023) https://doi.org/10.1088/1748-0221/18/10/P10017 arXiv:2302.11594 [physics.ins-det]

    Diefenbacher, S., Eren, E., Gaede, F., Kasieczka, G., Krause, C., Shekhzadeh, I., Shih, D.: L2LFlows: generat- ing high-fidelity 3D calorimeter images. JINST 18(10), 10017 (2023) https://doi.org/10.1088/1748-0221/18/10/P10017 arXiv:2302.11594 [physics.ins-det]

    Pith/arXiv arXiv 2023

  31. [31]

    JINST 19(02), 02003 (2024) https://doi.org/10.1088/1748-0221/19/02/P02003 arXiv:2303.10148 [hep-ex]

    Xu, A., Han, S., Ju, X., Wang, H.: Gen- erative machine learning for detector response modeling with a conditional nor- malizing flow. JINST 19(02), 02003 (2024) https://doi.org/10.1088/1748-0221/19/02/P02003 arXiv:2303.10148 [hep-ex]

    Pith/arXiv arXiv 2024

  32. [32]

    Buckley, M.R., Krause, C., Pang, I., Shih, D.: Inductive simulation of calorime- ter showers with normalizing flows. Phys. Rev. D 109(3), 033006 (2024) https://doi.org/10.1103/PhysRevD.109.033006 arXiv:2305.11934 [physics.ins-det]

    Pith/arXiv arXiv 2024

  33. [33]

    Pang, I., Shih, D., Raine, J.A.: Calorimeter shower superresolution. Phys. Rev. D 109(9), 092009 (2024) https://doi.org/10.1103/PhysRevD.109.092009 arXiv:2308.11700 [physics.ins-det]

    Pith/arXiv arXiv 2024

  34. [34]

    SciPost Phys

    Ernst, F., Favaro, L., Krause, C., Plehn, T., Shih, D.: Normalizing Flows for High-Dimensional Detector Simu- lations. SciPost Phys. 18, 081 (2025) https://doi.org/10.21468/SciPostPhys.18.3.081 arXiv:2312.09290 [hep-ph]

    Pith/arXiv arXiv 2025

  35. [35]

    Schnake, S., Kr¨ ucker, D., Borras, K.: Calo- PointFlow II Generating Calorimeter Show- ers as Point Clouds (2024) arXiv:2403.15782 [physics.ins-det]

    Pith/arXiv arXiv 2024

  36. [36]

    Du, H., Krause, C., Mikuni, V., Nach- man, B., Pang, I., Shih, D.: Unifying simulation and inference with normalizing flows. Phys. Rev. D 111(7), 076004 (2025) https://doi.org/10.1103/PhysRevD.111.076004 arXiv:2404.18992 [hep-ph]

    Pith/arXiv arXiv 2025

  37. [37]

    JINST 19(09), 09003 (2024) https://doi.org/10.1088/1748-0221/19/09/P09003 arXiv:2405.20407 [physics.ins-det]

    Buss, T., Gaede, F., Kasieczka, G., Krause, C., Shih, D.: Convolutional L2LFlows: gen- erating accurate showers in highly granular calorimeters using convolutional normal- izing flows. JINST 19(09), 09003 (2024) https://doi.org/10.1088/1748-0221/19/09/P09003 arXiv:2405.20407 [physics.ins-det]

    Pith/arXiv arXiv 2024

  38. [38]

    JINST 20(07), 07007 (2025) https://doi.org/10.1088/1748-0221/20/07/P07007 arXiv:2501.05534 [hep-ph] 13

    Birk, J., Gaede, F., Hallin, A., Kasieczka, G., Mozzanica, M., Rose, H.: OmniJet- α C: learning point cloud calorimeter simulations using generative trans- formers. JINST 20(07), 07007 (2025) https://doi.org/10.1088/1748-0221/20/07/P07007 arXiv:2501.05534 [hep-ph] 13

    Pith/arXiv arXiv 2025

  39. [39]

    Mikuni, V., Nachman, B.: Score-based generative models for calorimeter shower sim- ulation. Phys. Rev. D 106(9), 092009 (2022) https://doi.org/10.1103/PhysRevD.106.092009 arXiv:2206.11898 [hep-ph]

    Pith/arXiv arXiv 2022

  40. [40]

    JINST 18(11), 11025 (2023) https://doi.org/10.1088/1748-0221/18/11/P11025 arXiv:2305.04847 [physics.ins-det]

    Buhmann, E., Diefenbacher, S., Eren, E., Gaede, F., Kasieczka, G., Korol, A., Korcari, W., Kr¨ uger, K., McKeown, P.: CaloClouds: fast geometry-independent highly-granular calorimeter simula- tion. JINST 18(11), 11025 (2023) https://doi.org/10.1088/1748-0221/18/11/P11025 arXiv:2305.04847 [physics.ins-det]

    Pith/arXiv arXiv 2023

  41. [41]

    JINST 19(05), 05003 (2024) https://doi.org/10.1088/1748-0221/19/05/P05003 arXiv:2307.04780 [cs.LG]

    Acosta, F.T., Mikuni, V., Nachman, B., Arratia, M., Karki, B., Milton, R., Karande, P., Angerami, A.: Comparison of point cloud and image-based models for calorimeter fast simulation. JINST 19(05), 05003 (2024) https://doi.org/10.1088/1748-0221/19/05/P05003 arXiv:2307.04780 [cs.LG]

    Pith/arXiv arXiv 2024

  42. [42]

    JINST 19(02), 02001 (2024) https://doi.org/10.1088/1748-0221/19/02/P02001 arXiv:2308.03847 [hep-ph]

    Mikuni, V., Nachman, B.: CaloScore v2: single-shot calorimeter shower simulation with diffusion mod- els. JINST 19(02), 02001 (2024) https://doi.org/10.1088/1748-0221/19/02/P02001 arXiv:2308.03847 [hep-ph]

    arXiv 2024

  43. [43]

    Amram, O., Pedro, K.: Denoising diffu- sion models with geometry adaptation for high fidelity calorimeter simulation. Phys. Rev. D 108(7), 072014 (2023) https://doi.org/10.1103/PhysRevD.108.072014 arXiv:2308.03876 [physics.ins-det]

    Pith/arXiv arXiv 2023

  44. [44]

    JINST 19(04), 04020 (2024) https://doi.org/10.1088/1748-0221/19/04/P04020 arXiv:2309.05704 [physics.ins-det]

    Buhmann, E., Gaede, F., Kasieczka, G., Korol, A., Korcari, W., Kr¨ uger, K., McKeown, P.: CaloClouds II: ultra-fast geometry-independent highly-granular calorimeter simula- tion. JINST 19(04), 04020 (2024) https://doi.org/10.1088/1748-0221/19/04/P04020 arXiv:2309.05704 [physics.ins-det]

    Pith/arXiv arXiv 2024

  45. [45]

    SciPost Phys

    Jiang, C., Qian, S., Qu, H.: Choose your diffusion: Efficient and flexible ways to accelerate the diffusion model in fast high energy physics simula- tion. SciPost Phys. 18(6), 195 (2025) https://doi.org/10.21468/SciPostPhys.18.6.195 arXiv:2401.13162 [physics.ins-det]

    Pith/arXiv arXiv 2025

  46. [46]

    Kobylianskii, D., Soybelman, N., Dreyer, E., Gross, E.: Graph-based diffusion model for fast shower generation in calorimeters with irregular geometry. Phys. Rev. D 110(7), 072003 (2024) https://doi.org/10.1103/PhysRevD.110.072003 arXiv:2402.11575 [hep-ex]

    Pith/arXiv arXiv 2024

  47. [47]

    Jiang, C., Qian, S., Qu, H.: BUFF: Boosted Decision Tree based Ultra-Fast Flow match- ing (2024) arXiv:2404.18219 [physics.ins-det]

    Pith/arXiv arXiv 2024

  48. [48]

    SciPost Phys

    Favaro, L., Ore, A., Schweitzer, S.P., Plehn, T.: CaloDREAM – Detector Response Emulation via Attentive flow Matching. SciPost Phys. 18, 088 (2025) https://doi.org/10.21468/SciPostPhys.18.3.088 arXiv:2405.09629 [hep-ph]

    Pith/arXiv arXiv 2025

  49. [49]

    SciPost Phys

    Brehmer, J., Bres´ o, V., Haan, P., Plehn, T., Qu, H., Spinner, J., Thaler, J.: A Lorentz-equivariant transformer for all of the LHC. SciPost Phys. 19(4), 108 (2025) https://doi.org/10.21468/SciPostPhys.19.4.108 arXiv:2411.00446 [hep-ph]

    arXiv 2025

  50. [50]

    Buss, T., Gaede, F., Kasieczka, G., Korol, A., Kr¨ uger, K., McKeown, P., Mozzanica, M.: CaloHadronic: a diffusion model for the generation of hadronic showers (2025) arXiv:2506.21720 [physics.ins-det]

    arXiv 2025

  51. [51]

    Raikwar, P., Zaborowska, A., McKeown, P., Cardoso, R., Piorczynski, M., Yeo, K.: A Generalisable Generative Model for Multi- Detector Calorimeter Simulation (2025) arXiv:2509.07700 [physics.ins-det]

    Pith/arXiv arXiv 2025

  52. [52]

    In: 2nd Euro- pean AI for Fundamental Physics Conference (2025)

    Favaro, L., Giammanco, A., Krause, C.: Fast, accurate, and precise detector simula- tion with vision transformers. In: 2nd Euro- pean AI for Fundamental Physics Conference (2025)

    2025

  53. [53]

    Buss, T., Day-Hall, H., Gaede, F., Kasieczka, G., Kr¨ uger, K., Korol, A., Madlener, T., McKeown, P., Mozzanica, M., Valente, L.: CaloClouds3: Ultra-Fast Geometry- Independent Highly-Granular Calorimeter Simulation (2025) arXiv:2511.01460 14 [physics.ins-det]

    arXiv 2025

  54. [54]

    Hashemi, B., Krause, C.: Deep generative models for detector sig- nature simulation: A taxonomic review. Rev. Phys. 12, 100092 (2024) https://doi.org/10.1016/j.revip.2024.100092 arXiv:2312.09597 [physics.ins-det]

    arXiv 2024

  55. [55]

    Technical report, CERN, Geneva (2025)

    Photon showers in the ATLAS fast calorime- ter simulation: A voxelized dataset with min- imized information loss and improved ML models. Technical report, CERN, Geneva (2025). https://cds.cern.ch/record/2942061

    arXiv 2025

  56. [56]

    Zurbano Fernandez, I., et al.: High-Luminosity Large Hadron Collider (HL-LHC): Technical design report 10/2020 (2020) https://doi.org/10.23731/CYRM-2020-0010

    doi:10.23731/cyrm-2020-0010 2020

  57. [57]

    Gessinger-Befurt, P., Salzburger, A., Niermann, J.: The Open Data Detector Tracking System. J. Phys. Conf. Ser. 2438(1), 012110 (2023) https://doi.org/10.1088/1742-6596/2438/1/012110

    work page doi:10.1088/1742-6596/2438/1/012110 2023

  58. [58]

    https://cds.cern.ch/record/2285585

    ATLAS Collaboration: ATLAS Inner Tracker Pixel Detector: Technical Design Report (2017). https://cds.cern.ch/record/2285585

    arXiv 2017

  59. [59]

    https://cds.cern.ch/record/2257755

    ATLAS Collaboration: ATLAS Inner Tracker Strip Detector: Technical Design Report (2017). https://cds.cern.ch/record/2257755

    arXiv 2017

  60. [60]

    : A Common Track- ing Software Project

    Ai, X., et al. : A Common Track- ing Software Project. Comput. Softw. Big Sci. 6(1), 8 (2022) https://doi.org/10.1007/s41781-021-00078-8 arXiv:2106.13593 [physics.ins-det]

    Pith/arXiv arXiv 2022

  61. [61]

    In: 31st International Conference on Neural Information Processing Systems (2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszko- reit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention Is All You Need. In: 31st International Conference on Neural Information Processing Systems (2017)

    2017

  62. [62]

    In: Proceedings of the Fourth ACM International Conference on AI in Finance, pp

    Nagy, P., Frey, S., Sapora, S., Li, K., Cali- nescu, A., Zohren, S., Foerster, J.: Generative ai for end-to-end limit order book modelling: A token-level autoregressive generative model of message flow using a deep state space net- work. In: Proceedings of the Fourth ACM International Conference on AI in Finance, pp. 91–99 (2023)

    2023

  63. [63]

    GitHub (2024)

    Karpathy, A.: nanoGPT: The simplest, fastest repository for training/finetun- ing medium-sized GPTs. GitHub (2024). https://github.com/karpathy/nanoGPT

    2024

  64. [64]

    In: The Seventh Inter- national Conference on Learning Representa- tions (2017)

    Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: The Seventh Inter- national Conference on Learning Representa- tions (2017)

    2017

  65. [65]

    Zenodo (2025)

    Novak, T.: SiliconAI. Zenodo (2025). https://doi.org/10.5281/zenodo.17568416

    work page doi:10.5281/zenodo.17568416 2025

  66. [66]

    Zenodo (2025)

    Novak, T.: SiliconAI Val- idator. Zenodo (2025). https://doi.org/10.5281/zenodo.17567586 15

    work page doi:10.5281/zenodo.17567586 2025