Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Vision transformers can emulate Geant4 for both electromagnetic and hadronic calorimeter showers across arbitrary geometries, in tens of milliseconds on one GPU, and transfer learning cuts training time in half.

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 12:03 UTC pith:IDA6ZYH6

load-bearing objection Solid extension of CaloDREAM to irregular geometries with honest evaluations, but the abstract overstates fidelity and the transfer-learning evidence is thinner than the 'universal' claim. the 4 major comments →

arxiv 2601.05289 v2 pith:IDA6ZYH6 submitted 2026-01-07 hep-ph cs.LGhep-exphysics.ins-det

A universal vision transformer for fast calorimeter simulations

Luigi Favaro , Andrea Giammanco , Claudius Krause This is my paper
classification hep-ph cs.LGhep-exphysics.ins-det
keywords fast calorimeter simulationvision transformerconditional flow matchinggenerative networkstransfer learningGeant4 emulationirregular geometriesdetector simulation
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.

Fast calorimeter simulation is a bottleneck at the LHC and future colliders. This paper shows that a vision transformer trained with conditional flow matching can generate both electromagnetic and hadronic showers that are statistically indistinguishable from Geant4 across multiple detectors and geometries, in tens of milliseconds on one GPU. It also shows that pretraining the transformer on a large multi-detector dataset and fine-tuning on a target geometry roughly halves training time and improves data efficiency. The authors present the first patch-based simulation of a full detector with both electromagnetic and hadronic calorimeters.

Core claim

The paper's central claim is that a single vision-transformer architecture, extended to arbitrary voxelized geometries via patching, can emulate the Geant4 detector response across electromagnetic and hadronic showers with statistically negligible deviations. On the LEMURS multi-detector dataset, five detectors with different granularities are covered by one shape network, and the same backbone fine-tunes to CaloChallenge datasets and a high-granular ILC calorimeter. Generation stays at 10–100 ms per shower on a single NVIDIA A100 GPU. Fine-tuning from the LEMURS pretraining reaches the accuracy of a from-scratch network in roughly half the iterations and improves generalization when trainin

What carries the argument

The central object is a patch-based vision transformer (ViT) shape network trained with conditional flow matching: a universal backbone of transformer blocks acts on patch embeddings of the voxelized shower, while detector-specific embedding and head layers are reinitialized or interpolated during fine-tuning. This split—universal backbone plus light detector-specific layers—is what lets one network handle irregular geometries and enables transfer learning, because the backbone is expected to encode general shower features such as sparsity and dynamic range.

Load-bearing premise

The transfer-learning savings rely on the assumption that showers from different detectors and particle types are already close in the Gaussian latent space, so a pretrained backbone only needs to learn a small shift; the paper states this as a posit without an ablation.

What would settle it

Fine-tune the LEMURS-pretrained backbone on a detector that has a substantially different material composition and voxel layout (for example, a liquid-argon calorimeter or an ECal with much finer lateral granularity) and compare the ResNet AUC after the same number of iterations against a from-scratch run. If the fine-tuned network does not reach from-scratch performance in approximately half the iterations, the claimed transfer-learning advantage does not hold.

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

If this is right

  • Geant4-quality shower generation at 10–100 ms per shower on a single GPU, versus hours of CPU simulation per event.
  • The same trained backbone transfers to new detectors and particle types, halving the iterations needed to match from-scratch quality.
  • Fine-tuned networks generalize better than from-scratch networks at fixed training-data size, reducing Geant4 data production needs.
  • Irregular and high-granular geometries, including a full detector with electromagnetic and hadronic sections, are handled without mapping to a regular grid.
  • Metric choices matter: ResNet AUC and FPD can disagree, so a faithful evaluation should use the most discriminative classifier.

Where Pith is reading between the lines

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

  • If the latent-space proximity assumption generalizes, a publicly released pretrained calorimeter backbone would let experiments fine-tune on their own detector geometry with a fraction of the current computational and data budget.
  • A natural stress test is to fine-tune on a detector with very different materials or segmentation (e.g., liquid argon or a much finer ECal) and measure whether the training-time savings persist; the paper only tested similar cylindrical and cartesian pad geometries.
  • The metric discrepancy between FPD and ResNet AUC suggests current benchmarks may overstate fidelity; a classifier-based 'distance to Geant4' on the same low-level voxels is a stricter target.
  • The same universal-backbone recipe could plausibly extend to other sparse, high-dynamic-range detector signals such as tracker hits or Cherenkov images, since the patching only needs a voxel-to-patch map.

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

4 major / 4 minor

Summary. The paper presents a vision-transformer (ViT) based fast calorimeter simulation architecture built upon CaloDREAM, extended to irregular geometries via flexible patching and a 3D positional embedding. The authors benchmark this 'CaloDREAM++' model on CaloChallenge datasets ds1/ds2/ds3, the multi-detector LEMURS dataset, and a down-sampled CaloHadronic dataset, reporting generation times of O(10–100) ms per shower on GPU. They further propose a transfer-learning strategy: pretraining on LEMURS or ds2 and fine-tuning on target detectors, claiming reduced training iterations, improved data efficiency, and in some cases better fidelity. The paper includes code and data releases.

Significance. If the claims are correct, the paper is a strong step toward a single ViT backbone usable for multiple detector geometries and particle types, with practical training-cost savings. The public code/data release, the use of external Geant4 references, and the broad evaluation across regular, irregular, and full-detector geometries are notable strengths. The transfer-learning results, particularly the ds2 data-efficiency scan, are valuable. However, the abstract and conclusions overstate the indistinguishability of generated showers, and the FPD numbers in Table 3 are inconsistent with the metric definition and Table 1, undermining some quantitative claims. The universality of the transfer-learning benefit is not yet demonstrated for genuinely different geometries.

major comments (4)
  1. [Abstract / Section 7] The abstract and conclusions state that generated showers are 'statistically indistinguishable from Geant4 in multiple evaluation metrics.' This is contradicted by the paper's own ResNet classifier AUCs: Table 1 shows ds2 AUC=0.683(9) and ds3=0.799(9); Table 3 shows FCCeeALLEGRO AUC=0.688(16); Fig. 5 shows ds1-pi low-level AUC=0.633(3). These are far from the 0.5 expected for indistinguishable samples. The text in Section 4 even says 'only the more advanced ResNet extracts mismodeled generated features.' The claim should be qualified to high-level features or explicitly acknowledge low-level discrepancies.
  2. [Table 3] The FPD column in Table 3 reports values near 1.0033 for both Geant4 and ViT-CFM across all detectors. Table 1 reports FPD multiplied by 10^3, with Geant4 ds2 at 10.7(8), i.e. FPD ~0.0107. A distance metric between a sample and itself should be near zero, not ~1.003. These numbers are inconsistent and make the statement that 'the FPD metric scores all networks as Geant4-like' uninterpretable. Please correct the definition/units or explain the discrepancy before the FPD results can be used as evidence.
  3. [Section 6 / Figs. 7-9] The transfer-learning evidence is not sufficient for the claimed universality. Fine-tuning is demonstrated for LEMURS→ds2, where the target is the same Par04 geometry (the one-hot encoding is set to Par04SiW), and for ds2→ds3, a superresolution of the same geometry. The only geometrically different target, CaloHadronic, shows a small high-level AUC improvement (0.614(3)→0.600(4)) and no data-efficiency scan is provided. Section 2.2 posits that general features are encoded in the backbone and that fine-tuning needs only 'a smaller shift' in latent space, but this premise is not tested. The abstract's 'higher data efficiency' claim is therefore only demonstrated for near-identical detectors. Please either provide a data-efficiency scan on a genuinely different geometry or temper the universality claim.
  4. [Figure 9 (right)] For ds3, the fine-tuned ResNet AUC (0.789(9)) is not significantly different from the from-scratch value (0.799(9)) — the difference is within uncertainties. The paper acknowledges 'The high-granular ds3 shows a smaller gain,' but the abstract's phrase 'or altogether improves the fidelity of generated showers' is not supported by this result. The only significant fidelity improvement appears in ds2, so the wording should be softened accordingly.
minor comments (4)
  1. [Section 6] The claim that 'a third detector with the Par04 geometry can be encoded as a weighted combination of the two materials already seen during training' is speculative and not tested. Please mark it as a hypothesis or remove it unless an experiment supports it.
  2. [Section 4] The sentence 'The CaloDREAM++ samples for ds2 are statistically indistinguishable from the Geant4 reference' is directly contradicted by the ResNet AUC 0.683(9) in Table 1. Rephrase to refer to high-level features or to the selected metrics.
  3. [Table 3 caption] The caption says 'from a total of 200k showers.' Clarify whether this is per detector or across all detectors, and whether the FPD is computed on the same 200k sample as the classifiers.
  4. [Eq. (5)] The target velocity for the linear trajectory x(t)=(1-t)ϵ + t x0 is x0−ϵ; the notation (x−ϵ) in Eq. (5) is clear only after the expectation over x∼pdata is understood. Consider writing (x0−ϵ) explicitly for readability.

Circularity Check

0 steps flagged

No circularity: results are benchmarked against external Geant4 data and transfer-learning gains are measured, not assumed.

full rationale

The paper's central derivation chain is: define a conditional flow matching objective (Eq. 5), factor generation into an energy network and a ViT shape network (Eq. 7), train on public Geant4-derived datasets, and evaluate generated showers against held-out Geant4 reference samples using classifiers and FPD. No predicted quantity is also a fitted input: the energy ratios u and normalized voxels x are sampled from the trained densities, not read off the training data. The transfer-learning claims are empirical: fine-tuned and from-scratch networks are trained under the same budgets and compared via independently trained classifiers and FPD scores (Figs. 7-9). The latent-space proximity argument is explicitly introduced as 'We posit that general features, e.g the sparsity, of calorimeter showers are encoded in the large ViT backbone during pretraining' and is not used to define the evaluation metrics; the conclusions rest on the measured curves rather than on that posit. Self-citations to CaloDREAM [56] and the companion paper [57] provide an architectural starting point and a speed/accuracy variant, but they are not invoked as evidence for the present benchmarks, which are external (Geant4 reference data, CaloChallenge, LEMURS, CaloHadronic). No uniqueness theorem, no fitted-parameter-renamed-as-prediction, and no definitional identity between inputs and outputs was found. The only mild methodological note is that hyperparameters and patch sizes are chosen on the same datasets used for evaluation, but this is standard training/procedure and not a derivation-level circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The contributions rest on standard ML building blocks (CFM, ViT) and on hand-chosen geometric preprocessing choices (patch sizes, thresholds, downsampling, dummy bins) rather than on new physics entities. The transfer-learning premise and the sufficiency of the evaluation metrics are the largest unproven inputs.

free parameters (5)
  • Patch sizes per dataset = ds1 (1,1,5); ds2 (3,16,1); ds3 (3,10,3); LEMURS (3,16,1); CaloHad ECal (5,5,3), HCal (3,5,5)
    Chosen by hand (Table A3) to make each geometry patchifiable and to control the number of transformer patches; the efficiency and fidelity claims depend on them.
  • Sparsity/readout threshold x_th = 15.15 keV (CaloChallenge ds2/3), 1 keV (CaloHadronic)
    Hand-set thresholds used to define sparsity and to filter generated energy deposits; affects all sparsity and occupancy metrics.
  • CaloHadronic ECal downsampling = (180,180,30)->(15,15,10) via sum-pooling
    The task is simplified by coarsening the electromagnetic calorimeter; the claim of simulating the 'full detector' therefore rests on a reduced-resolution ECal.
  • Dummy zero-energy bins for ds1 = 4 extra bins per radial voxel in layers with a single angular bin
    Ad hoc addition to make layers patchifiable with patch size 5; the paper argues they are unbiased because they sit inside one patch, but they alter the geometry.
  • Weight decay schedule for small datasets = increased as training dataset size diminishes
    Tuned to avoid overfitting in the dataset-size scan (Fig. 9); a hyperparameter adjusted to the study.
axioms (5)
  • standard math Conditional flow matching with linear trajectories maps data to a standard Gaussian latent space.
    Used in Section 2.1; underlying theory of CFM assumed correct.
  • domain assumption Pre-processed calorimeter showers from different detectors and particles share enough structure (sparsity, dynamic range, spatial correlations) that a pretrained ViT backbone needs only a small distributional shift when fine-tuned.
    The transfer-learning claim in Section 6 rests on this; the paper states it as an observation, not a proven theorem.
  • domain assumption Classifier AUC and FPD on the chosen high-level observables are sufficient to certify that generated showers reproduce Geant4 for physics analyses.
    Used throughout Sections 3-5; if these metrics miss physics-relevant mismodelings, the fidelity claims overstate safety.
  • domain assumption Geant4 is an acceptable ground truth for calorimeter showers.
    Standard in the field; the paper compares to Geant4 samples from public datasets.
  • ad hoc to paper One-hot detector encoding can represent unseen detector configurations as a weighted combination of seen detectors (e.g., a third Par04-like material).
    Proposed in Section 6 as a way to inject prior knowledge; no validation shown for a truly new detector material.

pith-pipeline@v1.3.0-alltime-deepseek · 27566 in / 13705 out tokens · 125202 ms · 2026-08-03T12:03:38.674787+00:00 · methodology

0 comments
read the original abstract

The high-dimensional complex nature of detectors makes fast calorimeter simulations a prime application for modern generative machine learning. Vision transformers (ViTs) can emulate the Geant4 response with unmatched accuracy and are not limited to regular geometries. Starting from the CaloDREAM architecture, we demonstrate the robustness and scalability of ViTs on regular and irregular geometries, and multiple detectors. Our results show that ViTs generate electromagnetic and hadronic showers with minimal deviations from Geant4 in multiple evaluation metrics, while maintaining the generation time in the $\mathcal{O}(10-100)$ ms on a single GPU. Furthermore, we show that pretraining on a large dataset and fine-tuning on the target geometry leads to reduced training costs and higher data efficiency, or altogether improves the fidelity of generated showers.

Figures

Figures reproduced from arXiv: 2601.05289 by Andrea Giammanco, Claudius Krause, Luigi Favaro.

Figure 1
Figure 1. Figure 1: Schematic diagram of the vision transformer [59], which highlights the detector-specific and the universal part of the architecture. The color coded detector￾specific steps (see text for more details) indicate the components which may be reinitialized during fine-tuning. The universal ViT block only contains learnable transformations at patch-level objects. These weights, trained on a large corpus of data,… view at source ↗
Figure 2
Figure 2. Figure 2: Summary of the sliced evaluation for Einc ∈ [0.33 · Ei-max, 0.66 · Ei-max) GeV, θ ∈ [1.52, 1.62)), where Ei-max is 100 GeV for the FCC detectors and 1TeV for the Par04 and ODD detectors. We show the visible energy, the energy profile in the z direction, and the energy profile in the radial direction: (top) Par04SiW, (middle) ODD, and (bottom) FCCeeALLEGRO detectors. with a voxelized representation of a cal… view at source ↗
Figure 3
Figure 3. Figure 3: Summary of the sliced evaluation for Einc ∈ [0.66 · Ei-max, Ei-max) TeV, θ ∈ [2.1, 2.2)), where Ei-max is 100 GeV for the FCC detectors and 1TeV for the Par04 and ODD detectors. We show the visible energy, the energy profile in the z direction, and the energy profile in the radial direction: (top) Par04SiW, (middle) ODD, and (bottom) FCCeeALLEGRO detectors. Collider [71], namely the CLIC-like detector [72]… view at source ↗
Figure 4
Figure 4. Figure 4: Summary of the evaluation on the CaloChallenge-ds1-γ dataset. We show the center of energy and the shower width in layer-1, the AUC scores of a low- and high-level neural classifier, and the generation time on CPU, with batch size 1, and GPU, with batch size 100. 10−3 10−2 10−1 a.u. π + ds-1 Geant4 ViT-CFM 0.7 1.0 1.3 Model Geant4 −100 −50 0 50 100 hηi1 [mm] 0.1 1.0 10.0 δ[%] 10−3 10−2 10−1 a.u. π + ds-1 G… view at source ↗
Figure 5
Figure 5. Figure 5: Summary of the evaluation on the CaloChallenge-ds1-π + dataset. We show the center of energy and the shower width in layer-1, the AUC scores of a low- and high-level neural classifier, and the generation time on CPU, with batch size 1, and GPU, with batch size 100. improve over all the other submissions to the CaloChallenge. The lower dimensionality of the dataset implies that a smaller number of patches i… view at source ↗
Figure 6
Figure 6. Figure 6: Summary of the evaluation on the CaloHadronic dataset. We show the center of energy in the three coordinates x,y, and z for the entire calorimeter, the total visible energy Etot, and the average occupancy ⟨λ⟩ in the detector. In the table, we report the AUC score of a neural classifier and the generation time on CPU and GPU. 6 Transfer learning We present two clear cases where fine-tuning can be used to re… view at source ↗
Figure 7
Figure 7. Figure 7: Efficiency gain of fine-tuned networks for varying training iterations. The ResNet classifier AUC score of fine-tuned networks is significantly lower than that of the networks trained from scratch. (left) Pretraining on the LEMURS dataset, fine-tuned on the CaloChallenge-ds2. (right) Pretraining on the CaloChallenge-ds2, fine-tuned on the CaloChallenge-ds3. 100k 200k 300k 400k training iterations 0 100 200… view at source ↗
Figure 8
Figure 8. Figure 8: Efficiency gain of fine-tuned networks for varying training iterations. The FPD score of fine-tuned networks converges to the best value observed from a complete training in significantly less iterations than the networks trained from scratch. Notice that the the ResNet classifier in [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: (left) Efficiency gain of fine-tuned networks for varying dataset sizes. fine￾tuned networks better generalize as demonstrated by a ResNet classifier always trained with a total of 200k showers. Pretraining on the LEMURS dataset, fine-tuning on the CaloChallenge-ds2. (right) Summary table of classifier AUC scores after fine-tuning on the entire training dataset for the corresponding detectors. particular, … view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. CaloArt: Large-Patch x-Prediction Diffusion Transformers for High-Granularity Calorimeter Shower Generation

    physics.ins-det 2026-05 unverdicted novelty 5.0

    CaloArt achieves top FPD, high-level, and classifier metrics on CaloChallenge datasets 2 and 3 while keeping single-GPU generation at 9-11 ms per shower by combining large-patch tokenization, x-prediction, and conditi...

Reference graph

Works this paper leans on

78 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    Albrecht Jet al.2025J. Phys. G52030501 (Preprint2408.03881)

    Pith/arXiv arXiv

  2. [2]

    Harris R CERN hits one exabyte of stored experimental data from the LHC URLhttps://home.cern/news/news/computing/ cern-hits-one-exabyte-stored-experimental-data-lhc

  3. [3]

    Badger Set al.2023SciPost Phys.14079 (Preprint2203.07460)

    Pith/arXiv arXiv

  4. [4]

    Shanahan Pet al.2022 (Preprint2209.07559)

    Pith/arXiv arXiv 2022

  5. [5]

    Cranmer K, Brehmer J and Louppe G 2020Proc. Nat. Acad. Sci.11730055–30062 (Preprint1911.01429)

    Pith/arXiv arXiv

  6. [6]

    Badger S and Bullock J 2020JHEP06114 (Preprint2002.07516)

    Pith/arXiv arXiv

  7. [7]

    Core8073 (Preprint2412.12069)

    Bahl H, Elmer N, Favaro L, Haußmann M, Plehn T and Winterhalder R 2025 SciPost Phys. Core8073 (Preprint2412.12069)

    arXiv 2025

  8. [8]

    Beccatini L, Maltoni F, Mattelaer O and Winterhalder R 2025 (Preprint2512. 11036)

    2025

  9. [9]

    Gao C, Höche S, Isaacson J, Krause C and Schulz H 2020Phys. Rev. D101076002 (Preprint2001.10028)

    Pith/arXiv arXiv

  10. [10]

    Bothmann E, Janßen T, Knobbe M, Schmale T and Schumann S 2020SciPost Phys.8069 (Preprint2001.05478)

    Pith/arXiv arXiv

  11. [11]

    Heimel T, Winterhalder R, Butter A, Isaacson J, Krause C, Maltoni F, Mattelaer O and Plehn T 2023SciPost Phys.15141 (Preprint2212.06172)

    Pith/arXiv arXiv

  12. [12]

    Heimel T, Huetsch N, Maltoni F, Mattelaer O, Plehn T and Winterhalder R 2024 SciPost Phys.17023 (Preprint2311.01548)

    Pith/arXiv arXiv 2024

  13. [13]

    Heimel T, Mattelaer O, Plehn T and Winterhalder R 2025SciPost Phys.18017 (Preprint2408.01486)

    Pith/arXiv arXiv

  14. [14]

    Di Sipio R, Faucci Giannelli M, Ketabchi Haghighat S and Palazzo S 2019JHEP 08110 (Preprint1903.02433)

    Pith/arXiv arXiv

  15. [15]

    Verheyen R 2022SciPost Phys.13047 (Preprint2205.01697)

    Pith/arXiv arXiv

  16. [16]

    Butter A, Plehn T and Winterhalder R 2019SciPost Phys.7075 (Preprint 1907.03764)

    Pith/arXiv arXiv 1907

  17. [17]

    Butter A, Heimel T, Hummerich S, Krebs T, Plehn T, Rousselot A and Vent S 2023SciPost Phys.14078 (Preprint2110.13632)

    Pith/arXiv arXiv

  18. [18]

    Quétant G, Raine J A, Leigh M, Sengupta D and Golling T 2024Phys. Rev. D 110076023 (Preprint2406.13074)

    Pith/arXiv arXiv

  19. [19]

    Leigh M, Sengupta D, Quétant G, Raine J A, Zoch K and Golling T 2024SciPost Phys.16018 (Preprint2303.05376)

    Pith/arXiv arXiv

  20. [20]

    Phys.12100092 (Preprint2312.09597)

    Hashemi B and Krause C 2024Rev. Phys.12100092 (Preprint2312.09597)

    Pith/arXiv arXiv

  21. [21]

    Amram Oet al.2025Rept. Prog. Phys.88116201 (Preprint2410.21611) BIBLIOGRAPHY34

    arXiv

  22. [22]

    ATLAS Collaboration 2022 ATLAS Software and Computing HL-LHC Roadmap Tech. rep. CERN Geneva URLhttps://cds.cern.ch/record/2802918

    arXiv 2022

  23. [23]

    CMS Offline Software and Computing 2022 CMS Phase-2 Computing Model: Update Document Tech. rep. CERN Geneva URLhttps://cds.cern.ch/record/ 2815292

    2022

  24. [24]

    sciencedirect.com/science/article/pii/S0168900203013688

    Agostinelli S, Allison J, Amako K, Apostolakis J, Araujo H, Arce P, Asai M, Axen D, Banerjee S, Barrand G, Behner F, Bellagamba L, Boudreau J, Broglia L, Brunengo A, Burkhardt H, Chauvie S, Chuma J, Chytracek R, Cooperman G, Cosmo G, Degtyarenko P and Dell’Acqua A 2003Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectromete...

  25. [25]

    Allison J, Amako K, Apostolakis J, Araujo H, Arce Dubois P, Asai M, Barrand G, Capra R, Chauvie S, Chytracek R, Cirrone G, Cooperman G, Cosmo G, Cuttone G, Daquino G, Donszelmann M, Dressel M, Folger G, Foppiano F, Generowicz J, Grichine V, Guatelli S, Gumplinger P, Heikkinen A, Hrivnacova I, Howard A, Incerti S, Ivanchenko V, Johnson T, Jones F, Koi T, K...

  26. [26]

    Allison J, Amako K, Apostolakis J, Arce P, Asai M, Aso T, Bagli E, Bagulya A, Banerjee S, Barrand G, Beck B, Bogdanov A, Brandt D, Brown J, Burkhardt H, Canal P, Cano-Ott D, Chauvie S, Cho K, Cirrone G, Cooperman G, Cortés-Giraldo M, Cosmo G, Cuttone G, Depaola G, Desorgher L, Dong X, Dotti A, Elvira V, FolgerG,FrancisZ,GaloyanA,GarnierL,GayerM,GenserK,Gr...

  27. [27]

    Matchev K T, Roman A and Shyamsundar P 2022SciPost Phys.12104 (Preprint 2002.06307)

    Pith/arXiv arXiv 2002

  28. [28]

    Butter A, Diefenbacher S, Kasieczka G, Nachman B and Plehn T 2021SciPost Phys.10139 (Preprint2008.06545)

    Pith/arXiv arXiv

  29. [29]

    Bieringer S, Butter A, Diefenbacher S, Eren E, Gaede F, Hundhausen D, Kasieczka G, Nachman B, Plehn T and Trabs M 2022JINST17P09028 (Preprint2202. 07352)

  30. [30]

    Watts S J and Crow L 2025Mach. Learn. Sci. Tech.6025046 (Preprint2412. 18041)

  31. [31]

    Bahl H, Diefenbacher S, Elmer N, Plehn T and Spinner J 2025 (Preprint2509. 08048)

    2025

  32. [32]

    ch/record/2858890?ln=de

    Vaselli F, Rizzi A, Cattafesta F and Cicconofri G (CMS) URLhttps://cds.cern. ch/record/2858890?ln=de

    arXiv

  33. [33]

    Krammer N (CMS) 2025PoSLHCP2024280

  34. [34]

    Mazurek M 2025 Machine learning in LHCb Simulation: From fast to flash13th Large Hadron Collider Physics Conference(Preprint2511.02020)

    arXiv 2025

  35. [35]

    2025 Photon showers in the ATLAS fast calorimeter simulation: A vox- elized dataset with minimized information loss and improved ML models Tech. rep. CERN Geneva all figures including auxiliary figures are avail- able at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL- SOFT-PUB-2025-003 URLhttps://cds.cern.ch/record/2942061

    arXiv 2025

  36. [36]

    Krause C and Shih D 2023Phys. Rev. D107113003 (Preprint2106.05285)

    Pith/arXiv arXiv

  37. [37]

    Golling T, Heinrich L, Kagan M, Klein S, Leigh M, Osadchy M and Raine J A 2024 Mach. Learn. Sci. Tech.5035074 (Preprint2401.13537)

    Pith/arXiv arXiv 2024

  38. [38]

    Birk J, Hallin A and Kasieczka G 2024Mach. Learn. Sci. Tech.5035031 (Preprint 2403.05618)

    Pith/arXiv arXiv

  39. [39]

    Mikuni V and Nachman B 2025Phys. Rev. D111L051504 (Preprint2404.16091)

    arXiv

  40. [40]

    Wildridge A J, Rodgers J P, Colbert E M, yao Y, Jung A W and Liu M 2024 Bumblebee: Foundation Model for Particle Physics DiscoveryPostponed: Machine Learning and the Physical Sciences: Workshop at NeurIPS 2024(Preprint 2412.07867)

    Pith/arXiv arXiv 2024

  41. [41]

    Birk J, Gaede F, Hallin A, Kasieczka G, Mozzanica M and Rose H 2025JINST20 P07007 (Preprint2501.05534)

    Pith/arXiv arXiv

  42. [42]

    Mikuni V and Nachman B 2025Phys. Rev. D111054015 (Preprint2502.14652)

    arXiv

  43. [43]

    Core8046 (Preprint2503.19165)

    Tani L, Pata J and Birk J 2025SciPost Phys. Core8046 (Preprint2503.19165)

    Pith/arXiv arXiv

  44. [44]

    Hallin A 2025 Foundation models for high-energy physics2nd European AI for Fundamental Physics Conference(Preprint2509.21434)

    arXiv 2025

  45. [45]

    Bhimji W, Harris C, Mikuni V and Nachman B 2025 (Preprint2510.24066) BIBLIOGRAPHY36

    arXiv 2025

  46. [46]

    Dreyer F A, Grabarczyk R and Monni P F 2022Eur. Phys. J. C82564 (Preprint 2203.06210)

    Pith/arXiv arXiv

  47. [47]

    Chappell A and Whitehead L H 2022Eur. Phys. J. C821099 (Preprint 2207.03139)

    Pith/arXiv arXiv

  48. [48]

    Beauchesne H, Chen Z E and Chiang C W 2024JHEP02138 (Preprint2312. 06152)

  49. [49]

    Mokhtar F, Pata J, Garcia D, Wulff E, Zhang M, Kagan M and Duarte J 2025 Phys. Rev. D111092015 (Preprint2503.00131)

    Pith/arXiv arXiv 2025

  50. [50]

    Bonilla J L, Graczyk K M, Ankowski A M, Banerjee R D, Kowal B E, Prasad H and Sobczyk J T 2025 (Preprint2508.12987)

    arXiv 2025

  51. [51]

    Raikwar P, Zaborowska A, McKeown P, Cardoso R, Piorczynski M and Yeo K 2025 (Preprint2509.07700)

    Pith/arXiv arXiv 2025

  52. [52]

    Gaede F, Kasieczka G and Valente L 2025 (Preprint2512.00187)

    arXiv 2025

  53. [53]

    McKeown P, Raikwar P and Zaborowska A 2025 (Preprint2509.05108)

    arXiv 2025

  54. [54]

    Lipman Y, Chen R T Q, Ben-Hamu H, Nickel M and Le M 2023 (Preprint 2210.02747) URLhttps://arxiv.org/abs/2210.02747

    Pith/arXiv arXiv 2023

  55. [55]

    Shaul N, Singer U, Chen R T Q, Le M, Thabet A, Pumarola A and Lipman Y 2024 Bespoke non-stationary solvers for fast sampling of diffusion and flow models (Preprint2403.01329) URLhttps://arxiv.org/abs/2403.01329

    Pith/arXiv arXiv 2024

  56. [56]

    Favaro L, Ore A, Schweitzer S P and Plehn T 2025SciPost Phys.18088 (Preprint 2405.09629)

    Pith/arXiv arXiv

  57. [57]

    Favaro L, Giammanco A and Krause C 2025 Fast, accurate, and precise detector simulation with vision transformers2nd European AI for Fundamental Physics Conference(Preprint2509.25169)

    arXiv 2025

  58. [58]

    Tancik M, Srinivasan P P, Mildenhall B, Fridovich-Keil S, Raghavan N, Singhal U, Ramamoorthi R, Barron J T and Ng R 2020 Fourier features let networks learn high frequency functions in low dimensional domains (Preprint2006.10739) URL https://arxiv.org/abs/2006.10739

    Pith/arXiv arXiv 2020

  59. [59]

    Peebles W and Xie S 2023 Scalable diffusion models with transformers (Preprint 2212.09748) URLhttps://arxiv.org/abs/2212.09748

    Pith/arXiv arXiv 2023

  60. [60]

    Giannelli M F, Kasieczka G, Krause C, Nachman B, Salamani D, Shih D and Zaborowska A 2023 Fast calorimeter simulation challenge 2022 - dataset 1 version 3https://doi.org/10.5281/zenodo.8099322URLhttps://doi.org/10.5281/ zenodo.8099322

    work page doi:10.5281/zenodo.8099322urlhttps://doi.org/10.5281/ 2023

  61. [61]

    Buss T, Gaede F, Kasieczka G, Korol A, Krüger K, McKeown P and Mozzanica M 2025 (Preprint2506.21720)

    arXiv 2025

  62. [62]

    Kansal R, Li A, Duarte J, Chernyavskaya N, Pierini M, Orzari B and Tomei T 2023Phys. Rev. D107076017 (Preprint2211.10295) BIBLIOGRAPHY37

    Pith/arXiv arXiv

  63. [63]

    Heusel M, Ramsauer H, Unterthiner T, Nessler B and Hochreiter S 2017 Advances in Neural Information Processing Systems30arXiv:1706.08500 (Preprint1706.08500) URLhttps://proceedings.neurips.cc/paper_files/ paper/2017/file/8a1d694707eb0fefe65871369074926d-Paper.pdf

    Pith/arXiv arXiv 2017

  64. [64]

    Bińkowski M, Sutherland D J, Arbel M and Gretton A 2018International Conference on Learning RepresentationsarXiv:1801.01401 (Preprint1801.01401) URLhttps://openreview.net/forum?id=r1lUOzWCW

    Pith/arXiv arXiv

  65. [65]

    Szegedy C, Vanhoucke V, Ioffe S, Shlens J and Wojna Z 2015arXiv e-prints arXiv:1512.00567 (Preprint1512.00567)

    Pith/arXiv arXiv

  66. [66]

    Kansal R, Pareja C, Hao Z and Duarte J 2023Journal of Open Source Software8 5789 URLhttps://joss.theoj.org/papers/10.21105/joss.05789

    work page doi:10.21105/joss.05789

  67. [67]

    Das R, Favaro L, Heimel T, Krause C, Plehn T and Shih D 2024SciPost Phys.16 031 (Preprint2305.16774)

    Pith/arXiv arXiv

  68. [68]

    Giannelli M F, Kasieczka G, Krause C, Nachman B, Salamani D, Shih D and Zaborowska A 2022 Fast calorimeter simulation challenge 2022 - dataset 2https: //doi.org/10.5281/zenodo.6366271URLhttps://doi.org/10.5281/zenodo. 6366271

    work page doi:10.5281/zenodo.6366271urlhttps://doi.org/10.5281/zenodo 2022

  69. [69]

    Giannelli M F, Kasieczka G, Krause C, Nachman B, Salamani D, Shih D and Zaborowska A 2022 Fast calorimeter simulation challenge 2022 - dataset 3https: //doi.org/10.5281/zenodo.6366324URLhttps://doi.org/10.5281/zenodo. 6366324

    work page doi:10.5281/zenodo.6366324urlhttps://doi.org/10.5281/zenodo 2022

  70. [70]

    Allaire C, Gessinger P, Hdrinka J, Kiehn M, Kimpel F, Niermann J, Salzburger A and Sevova S 2022 Opendatadetector URLhttps://doi.org/10.5281/zenodo. 6445359

    doi:10.5281/zenodo 2022

  71. [71]

    Abada Aet al.(FCC) 2019Eur. Phys. J. C79474

  72. [72]

    Bacchetta Net al.2019 (Preprint1911.12230)

    Pith/arXiv arXiv 2019

  73. [73]

    Mlynarikova M (ALLEGRO) 2025EPJ Web Conf.32000022

  74. [74]

    Open Source Softw.85789

    Kansal R, Pareja C, Hao Z and Duarte J 2023J. Open Source Softw.85789

  75. [75]

    Aad Get al.(ATLAS) 2022Comput. Softw. Big Sci.67 (Preprint2109.02551)

    Pith/arXiv arXiv

  76. [76]

    Aihara Het al.(ILC) 2019 (Preprint1901.09829)

    Pith/arXiv arXiv 2019

  77. [77]

    Ernst F, Favaro L, Krause C, Plehn T and Shih D 2025SciPost Phys.18081 (Preprint2312.09290)

    Pith/arXiv arXiv

  78. [78]

    Toledo-Marin J Qet al.2025npj Quantum Inf.11114 (Preprint2410.22870)

    Pith/arXiv arXiv