Pith. sign in

REVIEW 3 major objections 7 minor 7 cited by

CaloDiT-2 shows that a diffusion model pre-trained on several calorimeters can be adapted to an unseen detector with up to 25x less data and 20x less training time than training from scratch — the first published pre-trained model for calor

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

2026-08-04 21:48 UTC pith:EMWJQXN7

load-bearing objection Solid engineering contribution with a real pre-training/adaptation result; the mesh-to-readout gap is acknowledged and is the main thing to watch. the 3 major comments →

arxiv 2509.07700 v1 pith:EMWJQXN7 submitted 2025-09-09 physics.ins-det hep-exphysics.comp-ph

A Generalisable Generative Model for Multi-Detector Calorimeter Simulation

Piyush Raikwar (CERN) , Anna Zaborowska (CERN) , Peter McKeown (CERN) , Renato Cardoso (CERN) , Mikolaj Piorczynski (CERN , Warsaw University of Technology) , Kyongmin Yeo (Thomas J. Watson Research Center) This is my paper
classification physics.ins-det hep-exphysics.comp-ph PACS 29.40.Vj
keywords calorimeter shower simulationgenerative diffusion modeltransformer architecturepre-training and adaptationfast simulationdetector-agnostic representationGeant4particle showers
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.

The paper aims to establish that fast simulation of calorimeter showers — normally a per-detector modelling effort requiring large training datasets — can be turned into a pre-train-once, adapt-anywhere problem. CaloDiT-2 is a transformer-based diffusion model that learns shower development on a detector-agnostic virtual mesh, pre-trained on four electromagnetic calorimeters. The central quantitative claim is that adapting this model to a fifth, unseen detector reproduces the accuracy of from-scratch training while using up to 25x less data and 20x less training time: 1,000 new samples and 5,000 steps match 25,000 samples and 50,000 steps. If true, this would cut the cost of producing simulation models for detectors under design — where geometry changes frequently — from days to hours, and the released model runs inside the Geant4 toolkit, where it is already shipped.

Core claim

CaloDiT-2 is a diffusion transformer that denoises a 9 x 16 x 45 shower image, conditioned on the photon's energy and angles plus a one-hot 'geometry condition' naming the detector. Showers are scored on a virtual cylindrical mesh (the Universal Representation of Showers from the Geant4 Par04 example), so one format serves several electromagnetic calorimeters. The claim: pre-training on four detectors teaches detector-agnostic shower physics, and adapting to a fifth unseen detector matches from-scratch accuracy with 25x less data (1,000 vs 25,000 showers) and 20x less training time — the first published pre-trained, adaptable model for shower simulation, shipped in Geant4, with a distilled s

What carries the argument

The central object is the Universal Representation of Showers: a virtual cylindrical mesh with 9 x 16 x 45 voxels in r, phi, z, finer than any detector readout. Scoring showers this way detaches data from specific geometries, so one model can ingest showers from many detectors. Three mechanisms carry the argument: the EDM continuous-time diffusion framework; consistency distillation, which compresses the teacher diffusion model into a single-step student; and a categorical geometry condition — a one-hot vector with one slot per pre-training detector plus an extra slot for the unseen target — giving adaptation an entry point that needs no geometric description of the new detector. The transfo

Load-bearing premise

The load-bearing premise is that the virtual mesh preserves all detector-relevant shower information, so accuracy measured on the mesh survives reintegration into detector-specific readouts; the paper itself flags this final validation as still to be done.

What would settle it

Take the adapted CaloDiT-2 for FCCeeALLEGRO, place generated showers back into the detector's real readout cells, and compare per-cell energy distributions and clustering observables against Geant4: if readout-level agreement is materially worse than mesh-level agreement, the adaptation gains would not survive into physics use. A simpler check: repeat the 1,000-shower adaptation on a detector close to the pre-training set and measure whether the 25x data advantage persists or shrinks.

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

If this is right

  • For an unseen detector, adaptation reaches from-scratch accuracy with up to 25x less training data and 20x less training time — in practice hours instead of days.
  • Detectors under active development, whose geometries change frequently, can regenerate their fast-simulation model after each design change rather than retraining from scratch.
  • The distilled single-step model generates a shower in roughly 100 ms on a single CPU core or about 3 ms on a GPU, with accuracy close to the 32-step diffusion model, making it usable inside Geant4.
  • Smaller experiments without the resources to design generative models can use the released pre-trained weights and adaptation scripts.
  • In the low-data regime (1,000 showers), adaptation stays generative while from-scratch training memorizes its training set, as shown by precision and density metrics.
  • All of these follow from the paper's own claims: the resource reduction, the Geant4 deployment, and the memorization comparison are reported results.

Where Pith is reading between the lines

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

  • If the geometry condition is a scalable slot, adding more detectors to the pre-training pool should keep improving adaptation quality and breadth — a path toward a genuine foundation model for calorimetry that the paper only gestures at.
  • The mesh's fixed 9 x 16 x 45 dimensionality caps which detectors can be represented; a point-cloud version trained under the same pre-train-then-adapt recipe is the natural next test, trading the clustering problem for sparsity.
  • Because adaptation with 1,000 samples avoids the memorization that from-scratch training exhibits, the pre-trained weights appear to encode physical priors about shower development; a testable corollary is that adaptation gains should scale with physical similarity between source detectors and target.
  • The choice to normalize jointly across detectors suggests the model learns a shared shower manifold; one could probe this by checking whether internal embeddings cluster by detector or by physical quantities such as radiation length.

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 / 7 minor

Summary. The manuscript presents CaloDiT-2, a transformer-based diffusion model for fast electromagnetic calorimeter shower simulation. The model uses an EDM diffusion framework, later distilled into a one-step consistency model. To enable cross-detector generalization, showers are scored on a virtual cylindrical mesh (9×16×45) rather than on each detector's native readout, and the model is conditioned on particle energy, angles, and a categorical detector identity. The authors train on four detectors from the new LEMURS dataset (Par04-SiW, Par04-SciPb, ODD, FCCeeCLD) and adapt to a fifth, FCCeeALLEGRO. On CaloChallenge Dataset-2 they report competitive FPD/KPD and fast inference. The central quantitative claim is that adaptation reaches the same FPD in 5,000 training steps with 1,000 new-detector showers as from-scratch training in 50,000 steps with 25,000 showers, a 20× training-time and 25× data reduction. The authors release datasets, pre-trained models, and a Geant4 integration. The paper repeatedly and honestly notes that all metrics are computed on the virtual mesh and that readout-level validation is not yet performed.

Significance. If the results hold, the contribution is significant: it is one of the first demonstrations of pre-training and adaptation for calorimeter shower simulation, with public datasets and models, and a plausible path to a foundation model for fast simulation. The release of LEMURS and the Geant4 integration are concrete community assets. The systematic comparison of adaptation versus from-scratch across data sizes and training steps is a useful benchmarking contribution, and the CaloChallenge results show a strong accuracy/speed trade-off. However, the headline resource-reduction claim is made for practical FastSim use, while all evidence is on the virtual mesh representation. The paper acknowledges this limitation in Section 2.3, but the abstract and conclusions present the 25×/20× savings without that qualification. The significance of the central claim therefore depends on whether the mesh-level gains survive reintegration into detector-specific readouts.

major comments (3)
  1. [Section 2.3 and Section 5.2 (Fig. 9)] The adaptation advantage is quantified exclusively on the 9×16×45 virtual mesh. Section 2.3 states: 'This validation would need to be taken one step further into the detector-specific readouts for each of the new adaptations.' This is load-bearing because the abstract and conclusion promise a FastSim model for detectors under development and a 25×/20× reduction in resources. The model is never exposed to readout segmentation, sampling fractions, inactive material, or cell-size non-uniformities; matching mesh observables does not guarantee correct projection onto a real readout. I request readout-level validation for at least FCCeeALLEGRO (e.g., placing generated showers back into the detector geometry and comparing readout-cell energy sums), or, failing that, a revision that explicitly limits the headline claims to virtual-mesh observables.
  2. [Section 5.2, Fig. 9] The comparison between adaptation and from-scratch training is confounded by unequal hyperparameters and lacks uncertainty quantification. The text says 'We empirically choose a learning rate of 0.001 for training from scratch and a learning rate of 0.0001 for adaptation,' and each plotted point is a single training run with no error bars. FPD is a relative measure, and the 50,000-sample estimate has statistical error. The strong statement that adaptation consistently outperforms training from scratch, and the quantitative 20×/25× claim, rest on point estimates under different optimization settings. Please provide repeated seeds or bootstrap confidence bands, and either match learning rates or show a small grid demonstrating that the conclusion is robust to learning-rate and EMA choices.
  3. [Section 4.1 (Figs. 5d, 6d) and Section 4.2] The total visible energy distribution is clearly not reproduced by either the EDM or CD model: the simulated distributions are visibly shifted relative to Geant4, as the text acknowledges ('except for the total visible energy'). This is a core shower observable and part of the CaloChallenge metrics. Yet the contributions state that CaloDiT-2 achieves 'good agreement with GEANT4 on widely used physics observables,' and Section 4.2 says it models 'nearly all shower observables defined in CaloChallenge with high accuracy.' This discrepancy needs to be quantified and either explained or removed before the broad accuracy claim is made.
minor comments (7)
  1. [Eq. (5) in Section 3.1.1] The loss is written in terms of s_phi(x,t), but the text says a separate network v_theta is trained and s_phi is only a preconditioned combination. Align the notation to avoid confusion.
  2. [Section 4.1] The text says '500 GeV e− fired' while the figures and dataset description refer to photons (γ). Clarify which particle is simulated.
  3. [Table 1] The KPD column header '×10^3' is ambiguous: values such as 0.089 for CaloDiT-2 EDM suggest the header should be '×10^-3' or that the values are in different units. Please check and correct.
  4. [References [56] and [60]] References [56] and [60] are the same paper ('Consistency Models'); consolidate to avoid duplicate citation.
  5. [Figure 7a] The caption does not define the axes or the Geant4 reference timing used for the speedup ratio. Include units and a one-sentence description of the calculation.
  6. [Section 5.3] The workflow says to condition on the '(K+1) detector position,' but the construction of the new one-hot entry is not specified in the main text. Clarify that the last entry is used for all unseen detectors.
  7. [Section 3.2.2] The low-energy threshold of 15.15 keV is introduced abruptly. State how it is chosen and whether it is detector-dependent.

Circularity Check

0 steps flagged

No significant circularity: the adaptation claim is validated on held-out Geant4 showers on a detector-agnostic mesh, and no fitted parameter is renamed as a prediction.

full rationale

The central claims — single-detector accuracy, CaloChallenge benchmarks, and cross-detector adaptation — are each evaluated against data not used in fitting. Section 5.2 computes FPD on 'a set of 50,000 test samples belonging to the entire distribution', i.e., held-out Geant4 showers from FCCeeALLEGRO, and compares adaptation to training-from-scratch under matched data and step budgets. The target detector is never included in the pre-training set; adaptation is genuine fine-tuning on new data. The one explicit limitation, noted in Section 2.3 ('This validation would need to be taken one step further into the detector-specific readouts for each of the new adaptations'), concerns whether mesh-level fidelity transfers to readout-level outputs; this is an external-validity risk, not circularity, because the model's mesh-level objective is independent of the readout. Self-references to the LEMURS dataset ([46], [47]) are provenance citations for the released data, and the CaloDiT-1/[28] reference is to an earlier model version; neither is invoked to justify the pre-training/adaptation result. No equation in the paper reduces a prediction to an input by construction; the FPD metric, mesh representation, and diffusion/consistency losses are all defined independently of the reported gains. Score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper's central claim depends on the virtual mesh representation and on several empirical hyperparameter choices. No new physical entities are introduced. The listed free parameters are the most consequential choices that could affect the adaptation comparison.

free parameters (5)
  • Adaptation learning rate = 0.0001
    Chosen empirically; affects how quickly the pre-trained model is fine-tuned and is a key knob in the adaptation comparison.
  • From-scratch learning rate = 0.001
    Chosen empirically for training from scratch; the different values between adaptation and from-scratch could bias the comparison.
  • Virtual mesh resolution = 9x16x45 (r x phi x z)
    Adopted from CaloChallenge Dataset-2; choice affects model capacity and fidelity.
  • Patch size = 3x2x3
    Found to offer trade-off between accuracy and efficiency; impacts tokenization.
  • Low-energy voxel threshold = 15.15 keV
    Voxels below this threshold are clipped to 0; affects voxel energy distribution.
axioms (4)
  • standard math Score-based diffusion (EDM) and consistency distillation frameworks from prior work are valid and applicable to this data modality.
    The method relies on the theoretical guarantees and practical recipes from Karras et al. and Song et al. (Sections 3.1.1, 3.1.2).
  • domain assumption The virtual cylindrical mesh captures shower information sufficient for downstream physics.
    Representation choice central to the detector-agnostic claim; the paper only validates on the mesh, not on readout-level observables (Section 2.1, 2.3).
  • domain assumption FPD/KPD on mesh-level observables is a meaningful proxy for physics fidelity.
    Used as the primary quantitative comparison in the adaptation study (Section 5.2).
  • domain assumption Uniform sampling of E, phi, theta provides unbiased training coverage.
    Dataset design choice, stated in Section 2.2, underlying the generalisation claim.

pith-pipeline@v1.3.0-alltime-deepseek · 144 in / 8330 out tokens · 134965 ms · 2026-08-04T21:48:47.843747+00:00 · methodology

0 comments
read the original abstract

Collider experiments, such as those at the Large Hadron Collider, use the Geant4 toolkit to simulate particle-detector interactions with high accuracy. However, these experiments increasingly require larger amounts of simulated data, leading to huge computing cost. Generative machine learning methods could offer much faster calorimeter shower simulations by directly emulating detector responses. In this work, we present CaloDiT-2, a diffusion model which uses transformer blocks. As is the case for other models explored for this task, it can be applied to specific geometries, however its true strength lies in its generalisation capabilities. Our approach allows pre-training on multiple detectors and rapid adaptation to new ones, which we demonstrate on the LEMURS dataset. It reduces the effort required to develop accurate models for novel detectors or detectors which are under development and have geometries that are changed frequently, requiring up to 25x less data and 20x less training time. To the best of our knowledge, this is the first pre-trained model to be published that allows adaptation in the context of particle shower simulations, with the model also included in the Geant4 toolkit. We also present results on benchmarks on Dataset-2 from the community-hosted CaloChallenge, showing that our models provide one of the best tradeoffs between accuracy and speed from the published models. Our contributions include a mechanism for the creation of detector-agnostic data representations, architectural modifications suitable for the data modality, a pre-training and adaptation strategy, and publicly released datasets and pre-trained models for broad use.

discussion (0)

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

Forward citations

Cited by 7 Pith papers

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

  1. Differentiable Surrogate for Detector Simulation and Design with Diffusion Models

    physics.ins-det 2026-01 unverdicted novelty 7.0

    A LoRA-adapted conditional diffusion surrogate for electromagnetic calorimeter showers matches key observables within 2% RMSE and reproduces directional trends in design-utility gradients.

  2. Lantern: Conflict-Aware Gradient Blending for Physics-Guided Diffusion Models in Calorimeter Simulation

    cs.LG 2026-07 conditional novelty 6.5

    GradBlend anchors diffusion updates to denoising while admitting physics auxiliaries, improving calorimeter shower FPD and CFD where PCGrad, GradNorm, IMTL-G, and ConFIG inflate FPD by 2–100×.

  3. CaloTrilogy: Toward a Breakthrough in One-Step, End-to-End, Physics-Guided Shower Generation for Modern Calorimeters

    hep-ex 2026-06 conditional novelty 6.0

    A one-step generative model for calorimeter showers, using MeanFlow, a learned Gaussian-mixture prior, and a physics-constrained loss, matches diffusion-model quality at far fewer evaluations.

  4. A universal vision transformer for fast calorimeter simulations

    hep-ph 2026-01 conditional novelty 6.0

    A vision-transformer flow-matching model generates calorimeter showers across regular and irregular detector geometries at millisecond speeds, and pretraining plus fine-tuning cuts training cost by about half.

  5. GPT-like transformer model for silicon tracking detector simulation

    physics.ins-det 2025-12 conditional novelty 6.0

    A decoder-only transformer trained on tokenized Geant4 hit sequences generates silicon tracker hits that reconstruct to near-Geant4-quality tracks for single muons.

  6. CaloTrilogy: Toward a Breakthrough in One-Step, End-to-End, Physics-Guided Shower Generation for Modern Calorimeters

    hep-ex 2026-06 unverdicted novelty 5.0

    Presents CaloTrilogy, a unified one-step generative model for high-granularity calorimeter showers that combines velocity field integration, learned priors, and physics losses to match SOTA quality.

  7. 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

72 extracted references · 34 canonical work pages · cited by 6 Pith papers

  1. [1]

    Geant4—a simulation toolkit

    Sea Agostinelli, John Allison, K al Amako, John Apostolakis, Henrique Araujo, Pedro Arce, Makoto Asai, D Axen, Swagato Banerjee, GJNI Barrand, et al. Geant4—a simulation toolkit. Nuclear instruments and methods in physics research section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 506(3):250–303, 2003

    2003

  2. [2]

    Technical report, CERN, Geneva, 2022

    ATLAS Software and Computing HL-LHC Roadmap. Technical report, CERN, Geneva, 2022

    2022

  3. [3]

    CMS Phase-2 Computing Model: Update Document

    CMS Offline Software and Computing. CMS Phase-2 Computing Model: Update Document. Technical report, CERN, Geneva, 2022

    2022

  4. [4]

    Calogan: Simulating 3d high energy particle showers in multilayer electromagnetic calorimeters with generative adversarial networks

    Michela Paganini, Luke de Oliveira, and Benjamin Nachman. Calogan: Simulating 3d high energy particle showers in multilayer electromagnetic calorimeters with generative adversarial networks. Physical Review D, 97(1):014021, 2018. 18

    2018

  5. [5]

    Fast and Accurate Simulation of Particle Detectors Using Generative Adversarial Networks

    Pasquale Musella and Francesco Pandolfi. Fast and Accurate Simulation of Particle Detectors Using Generative Adversarial Networks. Comput. Softw. Big Sci., 2(1):8, 2018

    2018

  6. [6]

    Precise simulation of electromagnetic calorimeter showers using a Wasserstein Generative Adversarial Network

    Martin Erdmann, Jonas Glombitza, and Thorben Quast. Precise simulation of electromagnetic calorimeter showers using a Wasserstein Generative Adversarial Network. Comput. Softw. Big Sci., 3(1):4, 2019

    2019

  7. [7]

    Fast simulation of a high granularity calorimeter by generative adversarial networks

    Gul Rukh Khattak, Sofia Vallecorsa, Federico Carminati, and Gul Muhammad Khan. Fast simulation of a high granularity calorimeter by generative adversarial networks. Eur. Phys. J. C, 82(4):386, 2022

    2022

  8. [8]

    Fast simulation of the ATLAS calorimeter system with Generative Adversarial Networks. 2020

    2020

  9. [9]

    Deep generative models for fast shower simulation in ATLAS

    The ATLAS collaboration. Deep generative models for fast shower simulation in ATLAS. Technical report, CERN, Geneva, 2018

    2018

  10. [10]

    Buhmann, S

    E. Buhmann, S. Diefenbacher, E. Eren, F. Gaede, Gregor Kasieczka, Anatolii Korol, and Katja Krüger. Getting high: High fidelity simulation of high granularity calorimeters with high speed. Comput. Softw. Big Sci., 5(1):13, 2021

    2021

  11. [11]

    Cresswell, Brendan Leigh Ross, Gabriel Loaiza-Ganem, Hum- berto Reyes-Gonzalez, Marco Letizia, and Anthony L

    Jesse C. Cresswell, Brendan Leigh Ross, Gabriel Loaiza-Ganem, Hum- berto Reyes-Gonzalez, Marco Letizia, and Anthony L. Caterini. Calo- Man: Fast generation of calorimeter showers with density estimation on learned manifolds. In 36th Conference on Neural Information Processing Systems: Workshop on Machine Learning and the Physical Sciences, 11 2022

    2022

  12. [12]

    New angles on fast calorimeter shower simulation.Mach

    Sascha Diefenbacher, Engin Eren, Frank Gaede, Gregor Kasieczka, Anatolii Korol, Katja Krüger, Peter McKeown, and Lennart Rustige. New angles on fast calorimeter shower simulation.Mach. Learn. Sci. Tech., 4(3):035044, 2023

    2023

  13. [13]

    Metahep: Meta learning for fast shower simulation of high energy physics experiments

    Dalila Salamani, Anna Zaborowska, and Witold Pokorski. Metahep: Meta learning for fast shower simulation of high energy physics experiments. Physics Letters B, 844:138079, 2023

    2023

  14. [14]

    Transformers for generalized fast shower simulation

    Raikwar, Piyush, Cardoso, Renato, Chernyavskaya, Nadezda, Jaruskova, Kristina, Poko- rski, Witold, Salamani, Dalila, Srivatsa, Mudhakar, Tsolaki, Kalliopi, Vallecorsa, Sofia, and Zaborowska, Anna. Transformers for generalized fast shower simulation. EPJ Web of Conf., 295:09039, 2024

    2024

  15. [15]

    Calo- vq: Vector-quantized two-stage generative model in calorimeter simulation

    Qibin Liu, Chase Shimmin, Xiulong Liu, Eli Shlizerman, Shu Li, and Shih-Chieh Hsu. Calo- vq: Vector-quantized two-stage generative model in calorimeter simulation. arXiv preprint arXiv:2405.06605, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Fast and accurate simulations of calorimeter showers with normalizing flows

    Claudius Krause and David Shih. Fast and accurate simulations of calorimeter showers with normalizing flows. Phys. Rev. D, 107(11):113003, 2023

    2023

  17. [17]

    Accelerating accurate simulations of calorimeter showers with normalizing flows and probability density distillation

    Claudius Krause and David Shih. Accelerating accurate simulations of calorimeter showers with normalizing flows and probability density distillation. Phys. Rev. D, 107(11):113004, 2023

    2023

  18. [18]

    L2LFlows: generating high-fidelity 3D calorimeter images

    Sascha Diefenbacher, Engin Eren, Frank Gaede, Gregor Kasieczka, Claudius Krause, Imahn Shekhzadeh, and David Shih. L2LFlows: generating high-fidelity 3D calorimeter images. JINST, 18(10):P10017, 2023

    2023

  19. [19]

    Normalizing Flows for High-Dimensional Detector Simulations

    Florian Ernst, Luigi Favaro, Claudius Krause, Tilman Plehn, and David Shih. Normalizing Flows for High-Dimensional Detector Simulations. SciPost Phys., 18:081, 2025

    work page 2025

  20. [20]

    Convolu- tional L2LFlows: generating accurate showers in highly granular calorimeters using convolu- tional normalizing flows

    Thorsten Buss, Frank Gaede, Gregor Kasieczka, Claudius Krause, and David Shih. Convolu- tional L2LFlows: generating accurate showers in highly granular calorimeters using convolu- tional normalizing flows. JINST, 19(09):P09003, 2024

    work page 2024

  21. [21]

    Score-based generative models for calorimeter shower simulation

    Vinicius Mikuni and Benjamin Nachman. Score-based generative models for calorimeter shower simulation. Phys. Rev. D, 106(9):092009, 2022

    work page 2022

  22. [22]

    CaloClouds: fast geometry- independent highly-granular calorimeter simulation

    Erik Buhmann, Sascha Diefenbacher, Engin Eren, Frank Gaede, Gregor Kasieczka, Anatolii Korol, William Korcari, Katja Krüger, and Peter McKeown. CaloClouds: fast geometry- independent highly-granular calorimeter simulation. JINST, 18(11):P11025, 2023. 19

    work page 2023

  23. [23]

    Denoising diffusion models with geometry adaptation for high fidelity calorimeter simulation

    Oz Amram and Kevin Pedro. Denoising diffusion models with geometry adaptation for high fidelity calorimeter simulation. Phys. Rev. D, 108(7):072014, 2023

    work page 2023

  24. [24]

    CaloScore v2: single-shot calorimeter shower simulation with diffusion models

    Vinicius Mikuni and Benjamin Nachman. CaloScore v2: single-shot calorimeter shower simulation with diffusion models. JINST, 19(02):P02001, 2024

    work page 2024

  25. [25]

    CaloClouds II: ultra-fast geometry-independent highly-granular calorime- ter simulation

    Erik Buhmann, Frank Gaede, Gregor Kasieczka, Anatolii Korol, William Korcari, Katja Krüger, and Peter McKeown. CaloClouds II: ultra-fast geometry-independent highly-granular calorime- ter simulation. JINST, 19(04):P04020, 2024

    work page 2024

  26. [26]

    CaloDREAM – Detector Response Emulation via Attentive flow Matching

    Luigi Favaro, Ayodele Ore, Sofia Palacios Schweitzer, and Tilman Plehn. CaloDREAM – Detector Response Emulation via Attentive flow Matching. SciPost Phys., 18:088, 2025

    work page 2025

  27. [27]

    CaloHadronic: a diffusion model for the generation of hadronic showers

    Thorsten Buss, Frank Gaede, Gregor Kasieczka, Anatolii Korol, Katja Krüger, Peter McKeown, and Martina Mozzanica. CaloHadronic: a diffusion model for the generation of hadronic showers. 6 2025

    work page 2025

  28. [28]

    Calochallenge 2022: A community challenge for fast calorimeter simulation

    Claudius Krause, Michele Faucci Giannelli, Gregor Kasieczka, Benjamin Nachman, Dalila Salamani, David Shih, Anna Zaborowska, Oz Amram, Kerstin Borras, Matthew R Buckley, et al. Calochallenge 2022: A community challenge for fast calorimeter simulation. arXiv preprint arXiv:2410.21611, 2024

    arXiv 2022

  29. [29]

    AtlFast3: The Next Generation of Fast Simulation in ATLAS

    Georges Aad et al. AtlFast3: The Next Generation of Fast Simulation in ATLAS. Comput. Softw. Big Sci., 6(1):7, 2022

    work page 2022

  30. [30]

    Par04 example

    GEANT4 Collaboration. Par04 example

    work page

  31. [31]

    Benedikt et al

    M. Benedikt et al. Future Circular Collider Feasibility Study Report: V olume 1, Physics, Experiments, Detectors. 4 2025

    work page 2025

  32. [32]

    Deep generative models for fast shower simulation in atlas

    Dalila Salamani, Stefan Gadatsch, Tobias Golling, Graeme Andrew Stewart, Aishik Ghosh, David Rousseau, Ahmed Hasib, and Jana Schaarschmidt. Deep generative models for fast shower simulation in atlas. In 2018 IEEE 14th International Conference on e-Science (e-Science), pages 348–348, 2018

    work page 2018

  33. [33]

    Omnijet-α: the first cross-task foundation model for particle physics

    Joschka Birk, Anna Hallin, and Gregor Kasieczka. Omnijet-α: the first cross-task foundation model for particle physics. Machine Learning: Science and Technology, 5(3):035031, 2024

    work page 2024

  34. [34]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018

    2018

  35. [35]

    Robust training of vector quantized bottleneck models

    Adrian Ła´ncucki, Jan Chorowski, Guillaume Sanchez, Ricard Marxer, Nanxin Chen, Hans JGA Dolfing, Sameer Khurana, Tanel Alumäe, and Antoine Laurent. Robust training of vector quantized bottleneck models. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1–7. IEEE, 2020

    work page 2020

  36. [36]

    Omnijet- αc: Learning point cloud calorimeter simulations using generative transformers

    Joschka Birk, Frank Gaede, Anna Hallin, Gregor Kasieczka, Martina Mozzanica, and Henning Rose. Omnijet- αc: Learning point cloud calorimeter simulations using generative transformers. arXiv preprint arXiv:2501.05534, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    On the opportunities and risks of foundation models

    Rishi Bommasani. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021

    Pith/arXiv arXiv 2021

  38. [38]

    A generalist agent

    Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022

    Pith/arXiv arXiv 2022

  39. [39]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  40. [40]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 20

    work page 2017

  41. [41]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  42. [42]

    Stable video diffusion: Scaling latent video diffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023

    Pith/arXiv arXiv 2023

  43. [43]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  44. [44]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022

    work page 2022

  45. [45]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  46. [46]

    Lemurs: Large-scale multi-detector electromagnetic universal representation of showers, September 2025

    Anna Zaborowska, Piyush Raikwar, and Peter McKeown. Lemurs: Large-scale multi-detector electromagnetic universal representation of showers, September 2025

    work page 2025

  47. [47]

    Lemurs dataset: Large-scale multi- detector electromagnetic universal representation of showers, 2025

    Peter McKeown, Piyush Raikwar, and Anna Zaborowska. Lemurs dataset: Large-scale multi- detector electromagnetic universal representation of showers, 2025

    work page 2025

  48. [48]

    Fast calorimeter simulation challenge 2022 - dataset 2, March 2022

    Michele Faucci Giannelli, Gregor Kasieczka, Claudius Krause, Ben Nachman, Dalila Salamani, David Shih, and Anna Zaborowska. Fast calorimeter simulation challenge 2022 - dataset 2, March 2022

    work page 2022

  49. [49]

    The Open Data Detector Tracking System

    Paul Gessinger-Befurt, Andreas Salzburger, and Joana Niermann. The Open Data Detector Tracking System. J. Phys. Conf. Ser., 2438(1):012110, 2023

    work page 2023

  50. [50]

    Bacchetta et al

    N. Bacchetta et al. CLD – A Detector Concept for the FCC-ee. 11 2019

    work page 2019

  51. [51]

    Design and performance of the calorimeter system for allegro fcc-ee detector concept

    Mlynarikova, Michaela. Design and performance of the calorimeter system for allegro fcc-ee detector concept. EPJ Web Conf., 320:00022, 2025

    work page 2025

  52. [52]

    Aidasoft/dd4hep, October 2018

    Markus Frank, Frank Gaede, Marko Petric, and Andre Sailer. Aidasoft/dd4hep, October 2018. webpage: http://dd4hep.cern.ch/

    work page 2018

  53. [53]

    Elucidating the design space of diffusion-based generative models, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models, 2022

    2022

  54. [54]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations, 2021

    2021

  55. [55]

    Interacting particle solutions of fokker–planck equations through gradient–log–density estimation

    Dimitra Maoutsa, Sebastian Reich, and Manfred Opper. Interacting particle solutions of fokker–planck equations through gradient–log–density estimation. Entropy, 22(8):802, July 2020

    work page 2020

  56. [56]

    Consistency models, 2023

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models, 2023

    work page 2023

  57. [57]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  58. [58]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  59. [59]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  60. [60]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 21

    work page 2023

  61. [61]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019

    2019

  62. [62]

    MiniCPM: Unveiling the potential of small language models with scalable training strategies

    Shengding Hu, Yuge Tu, Xu Han, Ganqu Cui, Chaoqun He, Weilin Zhao, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Xinrong Zhang, Zhen Leng Thai, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, dahai li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the potential of small language models with...

    work page 2024

  63. [63]

    Wang, David Leo Wright Hall, Percy Liang, and Tengyu Ma

    Kaiyue Wen, Zhiyuan Li, Jason S. Wang, David Leo Wright Hall, Percy Liang, and Tengyu Ma. Understanding warmup-stable-decay learning rates: A river valley loss landscape view. In The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  64. [64]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

    2019

  65. [65]

    Exponential moving average of weights in deep learning: Dynamics and benefits

    Daniel Morales-Brotons, Thijs V ogels, and Hadrien Hendrikx. Exponential moving average of weights in deep learning: Dynamics and benefits. arXiv preprint arXiv:2411.18704, 2024

    Pith/arXiv arXiv 2024

  66. [66]

    Bootstrap your own latent-a new approach to self-supervised learning

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271–21284, 2020

    2020

  67. [67]

    Level up your performance calculation of the fast shower simulation model

    Anna Zaborowska. Level up your performance calculation of the fast shower simulation model. https://indico.cern.ch/event/1253794/contributions/5588609/, 2023. ML4Jets 2023

    work page arXiv 2023

  68. [68]

    Evaluating generative models in high energy physics

    Raghav Kansal, Anni Li, Javier Duarte, Nadezda Chernyavskaya, Maurizio Pierini, Breno Orzari, and Thiago Tomei. Evaluating generative models in high energy physics. Physical Review D, 107(7):076017, 2023

    work page 2023

  69. [69]

    Improved precision and recall metric for assessing generative models

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019

    2019

  70. [70]

    Reliable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models. In International conference on machine learning, pages 7176–7185. PMLR, 2020

    work page 2020

  71. [71]

    CaloDiT-2

    GEANT4 Collaboration. CaloDiT-2

    work page

  72. [72]

    Aleksa, C

    M. Aleksa, C. Joram, P. Farthouat, A. Onnela, J. Blomer, C. Gargiulo, and P. Janot.Strategic R&D; Programme on Technologies for Future Experiments (No. CERN-OPEN-2018-006)., 2018. 22 Appendix A Additional Results On Single Detector 0 20 40 60 80 100 120Energy per event [MeV] Par04 E=50 GeV θ, φ=1.57, 0.0 rad Geant4 CaloDiT-2 EDM CaloDiT-2 CD 1 6 11 16 21 ...

    work page 2018